AI customer support that degrades gracefully instead of failing.

Backend lead and architect Jawad Ul Hadi built Omni.io, a multi-tenant RAG customer-support engine on NestJS, PostgreSQL with row-level security and pgvector, BullMQ, Gemini and React, designed so that every question still gets an answer when the AI layer fails. The system uses a three-tier fallback ladder — cited AI answer, then verbatim excerpts, then FAQ or human hand-off — plus a stored decision trace for each answer, and isolates tenants with Postgres row-level security under a NOBYPASSRLS role tested against a real database.

AI customer support that degrades gracefully instead of failing. Multi-tenant RAG on NestJS, PostgreSQL row-level security + pgvector, BullMQ, Gemini and React. Jawad Ul Hadi · Backend Lead & Architect · October 2026 Let's Connect https://gravatar.com/juhbukhari Repo https://github.com/JawadulHadi/omni-io Case study: Designing for AI Failure https://juh-bukhari.vercel.app/case-study Most RAG systems have one mode: working. If the embedding API is down, the model times out, or the model cites a passage it never saw, the customer gets either an error or a confident fabrication. Omni.io follows three rules: | Rule | How it is enforced | |---|---| | Every question gets an answer | A three-tier ladder in one service method that never throws: cited AI answer → verbatim excerpts → FAQ or human hand-off | | Every answer explains itself | A step-by-step decision trace, stored in the audit log and shown to support staff | | Tenants can't see each other's data, even through a bug | Postgres row-level security under a NOBYPASSRLS role, tested against a real database | flowchart LR subgraph People OP "Support team<br/ owner · admin · editor · viewer " VIS "Customer's website visitor<br/ anonymous " AIA "AI assistant<br/ Claude, Cursor, etc. " end subgraph Omni "Omni.io" SYS "Multi-tenant AI support engine" end subgraph External GEM "Google Gemini<br/ chat + embeddings" GOO "Google OAuth<br/ optional " end OP -- "Admin console<br/ GraphQL + WebSocket" -- SYS VIS -- "Embedded widget<br/ REST /w/:key" -- SYS AIA -- "MCP · Streamable HTTP<br/ personal access token" -- SYS SYS -- "embed · generate" -- GEM SYS -- "PKCE sign-in" -- GOO There are three ways in, each with its own trust level: | Channel | Caller | Authentication | What it can see | |---|---|---|---| | Console | Workspace members | 15-minute JWT + rotating refresh cookie | Everything their role allows | | Widget | Anonymous visitors | Public, rotatable widget key | Only documents and FAQs marked public | | MCP | AI assistants acting for a user | omni pat … personal access token | Exactly what that user can see | flowchart TB subgraph Edge "Edge · Caddy auto-HTTPS " CAD "Static console + widget.js<br/ reverse proxy" end subgraph App "Application tier" API "NestJS API<br/ GraphQL · REST · MCP · WebSocket<br/ stateless, horizontally scalable " WRK "Ingestion worker<br/ separate Node process<br/ chunk · embed · retention sweep" end subgraph Data "Data tier" PG "PostgreSQL 16 + pgvector 0.8<br/ RLS · HNSW · SECURITY DEFINER fns" RD "Redis 7<br/ BullMQ · rate limits · budgets · progress" end GEM "Gemini API" CAD -- "/graphql /auth/ /mcp /w/:key/ /health" -- API API -- "enqueue job ids only " -- RD RD -- "jobs" -- WRK WRK -. "progress events" .- RD RD -. "QueueEvents → GraphQL subscription" .- API API -- "omniio app role" -- PG WRK -- "omniio app role" -- PG API -- "query embedding + Tier 1 generation<br/ timeout + circuit breaker " -- GEM WRK -- "batch embeddings" -- GEM Why ingestion is queued but answering is synchronous. The split follows failure semantics, not threading: - Ingestion is long and retryable, and nobody is waiting on it. BullMQ gives durability, retries with backoff, and backpressure. A separate process means a 20 MB PDF can't starve the API of CPU, memory or DB connections. - Answering has a customer waiting. It runs inline under a hard time budget with circuit breakers. The worst case is a fast Tier 2 answer, never a queue. flowchart LR subgraph Entry "Entry points" GQL "GraphQL resolvers" REST "REST controllers" MCPC "MCP controller" end subgraph Features AUTH "AuthModule<br/ login · register · refresh · Google" TOK "ApiTokensModule<br/ MCP PATs" WS "WorkspacesModule<br/ members · roles · invites · ladder settings" DOC "DocumentsModule<br/ upload · paste · visibility · erasure" ING "IngestionModule<br/ producer · progress relay" ANS "AnswerModule<br/ resilience ladder" FAQ "FaqModule<br/ Tier 3 floor" WID "WidgetModule<br/ public key · theme · rotation" AUD "AuditModule<br/ answer.completed listener" MCP "McpModule<br/ 4 tools" HLT "HealthModule<br/ /health · systemInfo" end subgraph Infra "Shared infrastructure" DB "DbService<br/ tenant · withWorkspace · global " AI "AiProvider<br/ Gemini | Fake" RL "Redis rate limiter<br/ + in-process fallback" BS "BlobStorage<br/ none | local" end GQL -- AUTH & TOK & WS & DOC & ANS & FAQ & WID & AUD REST -- AUTH & DOC & WID & HLT MCPC -- MCP MCP -- WS & DOC & FAQ & ANS WID -- ANS ANS -- "emits answer.completed" -- AUD DOC -- ING ANS -- AI ANS & DOC & FAQ & WS & AUD & AUTH -- DB WID & AUTH & ANS -- RL DOC -- BS | Module | Main operations | |---|---| | AuthModule | POST /auth/login · /register · /refresh · /switch-workspace · /logout · /google | | ApiTokensModule | apiTokens , createApiToken , revokeApiToken | | WorkspacesModule | workspace , myWorkspaces , createInvitation , acceptInvitation , updateMemberRole , updateLadderSettings | | DocumentsModule | POST /documents/upload , documents , createDocumentFromText , deleteDocument | | IngestionModule | BullMQ producer, ingestionProgress subscription | | AnswerModule | askQuestion | | FaqModule | FAQ CRUD | | WidgetModule | GET /w/:key/config , POST /w/:key/ask , widgetConfig , rotateWidgetKey | | AuditModule | Internal listener; read-only answers | | McpModule | POST /mcp : list workspaces , list documents , list faqs , ask question | Every request passes the same gates in a fixed order. No feature code can skip them. php flowchart LR R "HTTP / WS request" -- TC "TenantContext middleware<br/ opens AsyncLocalStorage scope" TC -- CP "cookie-parser · JSON body 2 MB" CP -- AG{"AuthGuard<br/ @Public?"} AG -- "public route" -- RG AG -- "JWT valid<br/ PAT only on /mcp " -- FILL "fill context:<br/ userId · workspaceId" AG -- "invalid" -- X401 "401" FILL -- RG{"RolesGuard<br/ @Roles min vs<br/ LIVE membership row"} RG -- "insufficient" -- X403 "403" RG -- "ok" -- VP "ValidationPipe<br/ whitelist · forbidNonWhitelisted" VP -- H "Resolver / controller" H -- DBS "DbService.tenant <br/ BEGIN · set config app.workspace id, local · queries · COMMIT" H -. "any error" .- EF "AllExceptionsFilter<br/ sanitized, no stack traces" Roles are checked against the live workspace members row on every request, not against JWT claims. A demotion or removal takes effect on the next request, not when the token expires. One method, AnswerService.askQuestion , never throws. Every branch appends to the decision trace. php flowchart TD Q "Question ≤ 1,000 chars" -- RB{"Retrieval breaker open?"} RB -- "yes" -- T3 RB -- "no" -- E "Embed query + match chunks<br/ RETRIEVAL TIMEOUT MS = 4 s " E -- "error / timeout" -- T3 E -- F{"Any chunk ≥ workspace<br/ similarity floor?"} F -- "no · no relevant context" -- T3 F -- "yes" -- GB{"Generation breaker open?"} GB -- "yes · tier1 circuit open" -- T2 GB -- "no" -- BUD{"Daily model-call<br/ budget left?"} BUD -- "no · tier1 budget exhausted" -- T2 BUD -- "yes" -- G "Gemini · JSON schema output<br/ TIER1 TIMEOUT MS = 8 s " G -- "error · tier1 model error<br/ timeout · tier1 timeout" -- T2 G -- V1{"Valid JSON?"} V1 -- "no · tier1 invalid output" -- T2 V1 -- "yes" -- V2{"confidence ≥ threshold?"} V2 -- "no · tier1 low confidence" -- T2 V2 -- "yes" -- V3{"cites ≥ 1 passage, all<br/ ids ⊆ retrieved?"} V3 -- "no · tier1 no citations /<br/ tier1 invalid citation" -- T2 V3 -- "yes" -- V4{"grounding ≥ TIER1 MIN GROUNDING?<br/ answer words found in citations "} V4 -- "no · tier1 ungrounded" -- T2 V4 -- "yes" -- T1 "Tier 1 · cited AI answer" T2 "Tier 2 · top ≤ 3 excerpts verbatim" T3{"FAQ whole-word<br/ keyword match?"} -- "yes" -- F3 "Tier 3 · FAQ answer" T3 -- "no / DB error" -- H "Tier 3 · human hand-off message" T1 & T2 & F3 & H -- EV "emit answer.completed" EV -. "audit write fails → logged only" .- AUD "answers table" What "confident enough" means. Tier 1 needs four independent checks to pass: 1. Retrieval: at least one passage clears the per-workspace similarity floor default 0.6 . 2. Structure: the output is schema-valid JSON { answer, citedChunkIds, confidence } . 3. Provenance: every cited id was actually sent to the model. 4. Calibration + grounding: self-reported confidence clears the threshold, and enough of the answer's content words appear in the cited text. Self-reported confidence is poorly calibrated, so it is never the only gate. Cost honesty. Tokens spent on a Tier 1 attempt that was later rejected are still recorded, so the audit log reflects real spend. There is one breaker per dependency, so an embedding outage and a chat-model outage trip independently. php stateDiagram-v2 -- Closed Closed -- Open: 5 consecutive failures Open -- HalfOpen: 30 s cool-down elapsed HalfOpen -- Closed: the single probe succeeds HalfOpen -- Open: the probe fails note right of HalfOpen Exactly one probe request is let through, concurrent requests are short-circuited. end note sequenceDiagram autonumber actor U as Support agent participant C as Console urql participant A as NestJS API participant P as Postgres RLS participant G as Gemini participant L as AuditListener U- C: types question in Playground C- A: mutation askQuestion query A- A: AuthGuard → RolesGuard viewer → AskLimiter A- P: tenant tx: read ladder settings A- G: embed query 4 s budget G-- A: 768-dim vector A- P: tenant tx: match chunks ws, vec, k, public only=false P-- A: top-k chunks + similarity Note over A: no DB transaction is held during the model call A- G: generate JSON schema, delimited passages 8 s budget G-- A: { answer, citedChunkIds, confidence } A- A: validate JSON · citations ⊆ retrieved · confidence · grounding A-- C: AnswerOutcome { tier, answer, citations, trace } A-- L: event answer.completed L- P: tenant tx: insert into answers tokens, ids, trace, latency C-- U: tier badge + citations + decision trace sequenceDiagram autonumber actor E as Editor participant C as Console participant A as API participant X as PDF worker thread participant Q as Redis / BullMQ participant W as Ingestion worker participant G as Gemini participant P as Postgres E- C: drop file PDF / TXT / MD ≤ 20 MB C- A: POST /documents/upload A- A: ingest rate limit per workspace / hour opt PDF A- X: parse 30 s limit · 512 MB heap · max 2 concurrent X-- A: extracted text end A- P: insert document status = pending A- Q: add job { documentId, workspaceId } — ids only alt Redis down A- P: status = failed Retry button in console end Q- W: deliver job W- P: withWorkspace: load text W- W: chunk 1,200 chars / 200 overlap, word-aligned loop batches W- G: embed batch W-- Q: progress % Q-- A: QueueEvents A-- C: subscription ingestionProgress end W- P: upsert chunks by "documentId:chunkIndex" + delete stale tail W- P: status = ready Note over W: failures retry with backoff, UnrecoverableError → failed Idempotency. Chunk ids are deterministic ${documentId}:${chunkIndex} , so a retried job overwrites rather than duplicates. Stale tail chunks from a shorter re-ingest are deleted in the same transaction. Retention sweep. Every six hours the worker calls run retention . It deletes answer-audit rows older than ANSWER RETENTION DAYS default 90 , plus expired invitations, refresh tokens and API tokens. erDiagram USERS ||--o{ WORKSPACE MEMBERS : "belongs to" WORKSPACES ||--o{ WORKSPACE MEMBERS : has WORKSPACES ||--o{ DOCUMENTS : owns DOCUMENTS ||--o{ CHUNKS : "split into" WORKSPACES ||--o{ FAQS : owns WORKSPACES ||--o{ ANSWERS : audits WORKSPACES ||--o| WIDGET CONFIGS : themes WORKSPACES ||--o{ WORKSPACE INVITATIONS : issues USERS ||--o{ REFRESH TOKENS : "sessions" USERS ||--o{ API TOKENS : "MCP PATs" WORKSPACES { uuid id PK text name text plan text widget key "single source of truth" float confidence threshold float similarity floor } WORKSPACE MEMBERS { uuid workspace id FK uuid user id FK text role "owner|admin|editor|viewer" } USERS { uuid id PK text email text password hash "scrypt, not readable by app role" text google sub } DOCUMENTS { uuid id PK uuid workspace id FK text status "pending|processing|ready|failed" text source type text visibility "internal|public" text content text storage key } CHUNKS { text id PK "documentId:chunkIndex" uuid workspace id FK uuid document id FK vector embedding "768 dims, HNSW" text embedding model text content } FAQS { uuid id PK uuid workspace id FK text question text answer text array keywords "normalized" text visibility "internal|public" } ANSWERS { uuid id PK uuid workspace id FK text tier text channel "console|widget|mcp" text model int tokens in int tokens out text array retrieved chunk ids text array cited chunk ids float confidence float top similarity text decision note jsonb decision trace int latency ms } WIDGET CONFIGS { uuid workspace id FK jsonb theme timestamptz rotated at } WORKSPACE INVITATIONS { uuid id PK uuid workspace id FK text token hash "SHA-256" text role timestamptz expires at "7 days, single use" } REFRESH TOKENS { text token hash "SHA-256" uuid family id timestamptz used at timestamptz revoked at } API TOKENS { text token hash "SHA-256" text name timestamptz last used at timestamptz expires at timestamptz revoked at } Vector search. match chunks workspace id, query embedding, top k, public only runs server-side: - It filters by workspace id inside the SQL, on top of RLS. - It uses an HNSW index with hnsw.iterative scan = relaxed order . IVFFlat would post-filter, so small tenants would get fewer than k rows. - If the iterative scan still comes back short, it falls back to an exact search for that tenant only, which is cheap for small tenants. Migrations are plain, forward-only SQL files 0001 init → 0004 hardening , applied by a ~60-line runner under an advisory lock. Isolation is enforced in Postgres, in layers, so that no single application bug can cross tenants. flowchart TB L1 "â‘ Request context<br/ AuthGuard puts workspaceId in AsyncLocalStorage<br/ no workspace → DbService throws: fail closed " L2 "â‘¡ Unit of work<br/ one short transaction · set config 'app.workspace id', id, true <br/ transaction-local, can't leak to the next pooled request" L3 "â‘¢ Database role<br/ omniio app: not owner · not superuser · NOBYPASSRLS<br/ column grants hide users.password hash" L4 "â‘£ RLS policy on every tenant table<br/ USING / WITH CHECK workspace id = nullif current setting ... , '' ::uuid " L5 "⑤ Narrow SECURITY DEFINER functions<br/ auth find user · user workspaces · workspace role · resolve widget key<br/ invitation preview · accept invitation · run retention pinned search path " L6 "â‘¥ Proof<br/ e2e tests as omniio app: no-WHERE selects, cross-tenant insert/update/delete,<br/ vector search aimed at another tenant" L1 -- L2 -- L3 -- L4 L4 -.- L5 L4 -.- L6 Why nullif …, '' ? After a transaction-local set config , a pooled connection reads the setting back as '' , not NULL , and ''::uuid raises an error on every later unscoped query. Why no transaction across model calls? A slow LLM would pin pool connections. Each unit of work holds a connection for milliseconds. sequenceDiagram autonumber participant B as Browser tab participant A as API /auth participant P as Postgres B- A: POST /auth/login email, password A- P: auth find user — SECURITY DEFINER A- A: scrypt verify A- P: insert refresh token SHA-256, new family A-- B: access JWT 15 min, memory only + httpOnly SameSite=Strict cookie path /auth Note over B: access token expires B- B: Web Lock: one refresh across all tabs B- A: POST /auth/refresh cookie A- P: mark old token used · insert next token same family A-- B: new access JWT + rotated cookie alt an already-used token is presented theft A- P: revoke the whole family A-- B: 401 — every session in that family is signed out end | Concern | Decision | |---|---| | Access token | 15-minute JWT with typ: "access" , held in memory only | | Refresh token | Opaque, SHA-256 at rest, rotated on every use, reuse revokes the whole family | | Google sign-in | Authorization code + PKCE + state cookie. Never auto-linked to an existing password account by email, because sign-up doesn't verify email ownership | | Invitations | Single-use links valid for 7 days, bound to possession of the link. With ALLOW SIGNUP=false they are the only way to create an account | | Roles | owner admin editor viewer . Nobody can grant above their own role, only owners can modify owners, and there is always ≥ 1 owner | | Production boot guard | The API refuses to start with COOKIE SECURE=false or the example JWT SECRET | sequenceDiagram autonumber actor V as Visitor participant S as Customer site participant J as widget.js Shadow DOM participant A as API /w/:key participant R as Redis participant P as Postgres S- J: <script src=".../widget.js" data-omniio-key=KEY J- A: GET /w/KEY/config A- P: resolve widget key KEY alt key unknown or rotated A-- J: 404 → widget stays hidden else Postgres down A-- J: default theme end V- J: asks a question J- A: POST /w/KEY/ask CORS , no credentials A- R: rate limit per IP + per workspace A- A: ladder with public only = true · no trace returned A-- J: { tier, answer, citations } - Isolation from the host page: the widget renders in a Shadow DOM, ships as a separate ~72 kB gzipped IIFE bundle, and is ASCII-only so it works on pages without a UTF-8 charset. - Data exposure: it answers only from public documents and FAQs. Both are internal by default, so Tier 2 and Tier 3 can't leak internal content to anonymous visitors. - Key rotation: takes effect immediately and doesn't affect console sessions. sequenceDiagram autonumber participant M as MCP client participant A as POST /mcp participant T as McpToolsService participant P as Postgres M- A: JSON-RPC tools/call + Bearer omni pat … A- A: AuthGuard PAT accepted only on /mcp → user A- T: new stateless McpServer bound to user T- P: workspace role workspaceId, userId alt not a member T-- M: isError: "You are not a member of that workspace." else member T- P: withWorkspace workspaceId — same RLS scope as the console T-- M: JSON result end | Tool | Input | Effect | |---|---|---| | list workspaces | — | Read-only: the caller's memberships and roles | | list documents | workspaceId | Read-only: documents with ingestion status | | list faqs | workspaceId | Read-only: Tier 3 FAQ entries | | ask question | workspaceId , query ≤ 1,000 chars | Runs the ladder. It calls Gemini, writes an audit row and is rate-limited per user | Every tool declares a title, an input schema and all four MCP annotation hints readOnlyHint , destructiveHint , idempotentHint , openWorldHint , so clients can warn before a tool with side effects runs. The transport is Streamable HTTP in stateless mode HTTP+SSE is deprecated in the MCP spec , so there are no sessions to store or scale. OAuth 2.1 is on the roadmap. The reference deployment is one server running Docker Compose. The same images run unchanged on container platforms. php flowchart TB NET "Internet" -- ":80 / :443 HTTP/3 " -- WEB subgraph Host "Single Linux host · docker compose project: omniio " WEB "web · Caddy<br/ auto-HTTPS · HSTS · nosniff<br/ console + widget static files" API "api · node dist/main.js :3000<br/ healthcheck /health" WRK "worker · node dist/worker.js" MIG "migrate · node dist/migrate.js<br/ runs once, then exits " PG "postgres · pgvector/pgvector:pg16<br/ volume pgdata" RD "redis:7-alpine · AOF<br/ volume redisdata" end WEB -- "reverse proxy" -- API MIG -- "owner role DATABASE MIGRATOR URL " -- PG API -- "omniio app" -- PG WRK -- "omniio app" -- PG API -- RD WRK -- RD API & WRK -- "HTTPS" -- GEM "Gemini API" PG -. "healthy" .- MIG MIG -. "completed successfully" .- API & WRK API -. "healthy" .- WEB Start-up order: Postgres becomes healthy → migrate runs as the schema owner and exits 0 → the API and worker start as omniio app → Caddy starts once the API is healthy. Scaling path: | Component | Scales | Note | |---|---|---| | API | Horizontally | Stateless apart from in-process circuit breakers | | Worker | Horizontally | BullMQ distributes jobs; concurrency: 2 per process | | Postgres | Vertically + read replicas | Managed options need pgvector ≥ 0.8 for hnsw.iterative scan | | Redis | Managed instance | Holds only queues and counters, no source data | | Console / widget | CDN | Same-site with the API, because the refresh cookie is SameSite=Strict | | Dependency fails | Detection | Customer gets | Recorded as | |---|---|---|---| | Gemini chat model errors | exception | Tier 2 excerpts | tier1 model error | | Gemini chat model hangs | 8 s AbortSignal | Tier 2 excerpts | tier1 timeout | | Repeated model failures | breaker open | Tier 2, no model call | tier1 circuit open | | Model returns bad JSON | schema parse | Tier 2 | tier1 invalid output | | Model invents a citation | ids ⊄ retrieved | Tier 2 | tier1 invalid citation | | Answer not supported by its citations | grounding ratio | Tier 2 | tier1 ungrounded | | Daily budget spent | Redis counter | Tier 2 | tier1 budget exhausted | | Embedding API down or slow | exception / 4 s budget / breaker | Tier 3 FAQ or hand-off | retrieval failed / retrieval timeout / retrieval circuit open | | Nothing relevant indexed | similarity floor | Tier 3 | no relevant context | | Audit write fails | listener catch | Same answer | logged only | | Redis down | client error | Same answers; per-process rate limits; uploads saved as failed + Retry | /health → degraded | | Postgres down | client error | Widget: hand-off message + default theme | /health → 503 | | PDF bomb or slow parse | worker-thread limits | Upload rejected; API stays responsive | document error | | Layer | Stack | |---|---| | API | NestJS 12, GraphQL Apollo 5, code-first , REST, MCP SDK, zod, class-validator | | Data | PostgreSQL 16, pgvector 0.8 HNSW + iterative scan , RLS, plain SQL migrations | | Jobs | BullMQ 6 on Redis 7, separate worker process | | AI | Gemini gemini-3.5-flash + gemini-embedding-2 768 dims behind an AiProvider interface, plus a deterministic fake with failure injection | | Console | React 19, Vite 8, React Router 7, Tailwind v4, shadcn/ui, urql, graphql-ws | | Widget | Standalone React IIFE bundle in a Shadow DOM | | Delivery | Docker Node 24 LTS , Caddy, GitHub Actions typecheck, unit, RLS e2e on real pgvector, image builds , Dependabot | | ADR | Decision | |---|---| | 0001 https://github.com/JawadulHadi/omni-io/blob/main/docs/adr/0001-tenant-isolation-in-postgres.md | Tenant isolation in Postgres, not application code | | 0002 https://github.com/JawadulHadi/omni-io/blob/main/docs/adr/0002-short-tenant-transactions.md | One short tenant transaction per unit of work | | 0003 https://github.com/JawadulHadi/omni-io/blob/main/docs/adr/0003-plain-sql-migrations.md | Plain SQL migrations instead of an ORM | | 0004 https://github.com/JawadulHadi/omni-io/blob/main/docs/adr/0004-sync-answers-queued-ingestion.md | Synchronous answers, queued ingestion | | 0005 https://github.com/JawadulHadi/omni-io/blob/main/docs/adr/0005-model-output-is-untrusted.md | Model output is untrusted input | | 0006 https://github.com/JawadulHadi/omni-io/blob/main/docs/adr/0006-rotating-refresh-tokens.md | Rotating refresh tokens with reuse detection | | 0007 https://github.com/JawadulHadi/omni-io/blob/main/docs/adr/0007-urql-and-a-separate-widget-bundle.md | urql and a separate widget bundle | | 0008 https://github.com/JawadulHadi/omni-io/blob/main/docs/adr/0008-mcp-streamable-http.md | MCP over Streamable HTTP, stateless, under the caller's scope | Source code, ADRs and full documentation: Source https://github.com/JawadulHadi/omni-io · Case study: Designing for AI Failure https://juh-bukhari.vercel.app/case-study Designing for AI Failure