Shyam Sekhar

A self-hosted Kubernetes platform in two layers: a Redis-backed task engine (priority queues, retries, dead-letter, cost-aware autoscaling) and an AI agent on top. Describe an operation in plain English and the agent either answers it — a read-only, human-approved query that returns a result and writes nothing — or changes something: it drafts schema-checked code, waits for human approval, and opens a real GitHub pull request. It follows an authored runbook when one fits and plans the change itself from the schema when none does, either way through the same guardrails. It drafts; it never applies to production on its own.

• Distributed task engine (Go + Redis): priority queues (high/default/low) with weighted draining, exponential-backoff retries with jitter, a dead-letter queue, and at-least-once delivery backed by a crash-recovery reclaim loop. Load-tested at 728 enqueue requests/sec, 0% errors, ~10 ms median on a single replica.
• The AI agent runs as a resumable job: it matches your request to a runbook (embeddings shortlist → LLM re-rank by intent) or, when none fits, plans the change itself from the schema — down the identical pipeline of validation, policy, human approval, and PR-only. It drafts step by step and pauses at every approval gate — checkpointed, so a run survives a closed browser or a worker restart. Read-only "Ask" requests take the same gate but return an answer instead of a PR, and a change-shaped question typed in Ask is caught and offered a one-click switch to Change. Drafting runs on a per-request local (Ollama) or cloud (Groq) toggle with multi-key rotation and 429 failover.
• Grounded in real schema, guarded against drift: generated SQL is grounded in a curated, multi-database schema document rather than live introspection; a daily check plus an ArgoCD post-sync hook pause the agent the moment the database drifts from that document, until it is reconciled.
• Org policy, an eval loop, and ChatOps around the AI: every drafted SQL artifact is checked against org-authored rules before a PR can open — no irreversible DROP/TRUNCATE, GRANT only to allow-listed roles, migrations must carry a rollback, no unbounded UPDATE/DELETE — with violations surfaced at the approval gate and in the PR body. A regression suite gates content-repo CI on a model bump, a live reviewer-decision metric flags runbooks whose drafts keep getting rejected, and a pending gate pings Slack with a review deep link — auto-cancelling if left undecided past a configurable timeout.
• Full platform-engineering spine: k3s delivered via Helm + ArgoCD GitOps, Terraform, and CI; Prometheus / Grafana / Loki / OpenTelemetry for observability; a cost-aware daemon that scales idle workers to zero; secrets managed by the Infisical operator. Engine and content are fully separated — workflows and schema live in their own repo, pulled at runtime by a git-sync sidecar with in-process hot-reload, so content changes ship with no rebuild or redeploy.

An AI-native practice-management system and EMR for small clinics (≤5 doctors) and solo practitioners — capture, schedule, and follow up in one system. Every doctor belongs to a clinic account; live at medxcore.osforlife.in.

• AI clinical capture — Groq Whisper voice dictation, Llama-Vision OCR of paper prescriptions and lab reports, and manual entry, all funnelled through a human-in-the-loop review → confirm loop; auto after-visit summaries (English + Malayalam), pre-visit AI briefings, and a clinical safety-net.
• Live token queue (“where's my train”) over Postgres LISTEN/NOTIFY → SSE, with a predictive ETA that blends the doctor's set per-patient time with a learned rolling average; async AI pipeline on pg-boss (202-accept + poll).
• Security-first data model — Postgres Row-Level Security partitions clinical data per authoring doctor with patient-controlled, revocable cross-doctor sharing; layered RBAC staff profiles (Doctor · Nurse · Pharmacist · Front-office · Super-admin · Custom); append-only audit log on every clinical write; raw audio purged after transcription.
• Scheduling & lifecycle — month + day (operatory) calendar with conflict-aware booking and a same-family overlap exception, versioned medical history, and a recall / 6-month review engine on a pg-boss cron; doctor community profiles (media + posts).

A cross-platform iOS + Android app that turns long-term goals into daily action — and it is shipped, in real users' hands.

• Modular NestJS backend (Auth · Goals · Habits · Insights); JWT auth with refresh-token rotation (reuse detection → revoke session family) and per-device sessions in Redis.
• LLM goal decomposition (goal → weekly milestones → daily actions) with schema validation before persistence; automated weekly insight generation.
• PostgreSQL time-series for habits/goals — streak calculation via gap-and-island window functions, rolling completion-rate aggregations on a (user_id, date) index.
• React Native app with shared logic + platform-specific UI; push notifications. Shipped to production with GitHub Actions CI/CD.

A dashboard that scans NSE stocks and ETFs, generates swing-trade signals, paper-trades them, and backtests the whole strategy — end to end.

• Pipeline: yfinance OHLCV/fundamentals → EMA indicators → buy/sell/hold signals + fundamental score → portfolio tracker → paper-trading simulation → backtest engine.
• Two Docker services (Streamlit UI + a separate scheduler container, TZ Asia/Kolkata, persistent volume) — interactive vs batch separation.
• TTL cache shields the rate-limited data source from UI auto-refresh.

An online multiplayer Codenames clone with a 1,000-word Manglish (Malayalam) word bank — playable in real time with friends.

• No custom backend — the database is the trust boundary. Row-level security + a masked board view ensure operatives never receive secret card colors (hidden information enforced server-side in Postgres).
• Game moves are atomic Postgres RPCs (cheat-resistant); Supabase Realtime fans out state to all players.

A study platform that turns a syllabus-shaped notes library into a fast, searchable web app — in fact, the very site you are reading this on.

• Local embeddings (FastEmbed/ONNX, free/CPU) → semantic search; content + vectors synced to Supabase; the app reads content live.
• Render buildFilter separates content-publish cadence from code-deploy cadence — content commits never redeploy the app.
• CI validates content JSON on every push; password gate; secrets in the dashboard, never in the repo.

The operational control plane behind CareStack's multi-tenant, sharded, multi-region dental SaaS — one place to automate service requests, data-patch workflows, incident tracking, and monitoring.

• Self-healing data-patch pipeline: parameterised request → pull script from GitLab → substitute env vars → pre-run validation gate → peer/lead review or risk-proportional auto-merge → on merge, GitLab CI pushes a queue message + uploads the script to Azure Blob → an Argo job dequeues and executes against the correct shard via Flyway → live status streamed → post-execution validation → automatic MR revert + compensating script on failure.
• Led Sev1–Sev3 incident response; blameless post-mortems; integrated the Support User Access Tool.

An internal .NET Core ETL and migration platform that replaced the SSIS-based tooling used to migrate customers off competitor systems.

• Source-native stored procedures per engine behind a heterogeneous driver-abstraction layer.
• Staging hop (OneDB) decouples N sources from the target transform (N extractors + 1 transform vs N×M).
• Configurable scheduler, retry logic, execution audit; post-migration reconciliation (counts/checksums/invariants).

Completely removes a patient and all of their data across several polyglot microservice databases — without breaking integrity, tenant isolation, or analytics.

• An idempotent, ordered saga (no 2PC across heterogeneous engines); delete children-before-parents per store.
• Every delete scoped to tenant_id + patient_id (no blind cascades → co-tenant isolation).
• Deletion propagated into analytics/ETL for aggregate consistency; tombstones to prevent re-add races; reconciliation to prove completeness.

Centralized control over feature access and rollout eligibility, plus an internal portal for controlled feature-flag rollout, cohort segmentation, and A/B configuration across tenant groups.

• Multi-tenant shard routing, HIPAA-aware audit logging across microservice boundaries, tenant-provisioning automation.
• Phased/gradual rollouts and kill-switch feature flags.

A CLI that grants and revokes support-engineer database access across regions and database engines — safely, and all-or-nothing.

• Spans multiple regions (US/UK/AU/IE + sandboxes) and polyglot stores — sharded SQL Server PMS databases, MySQL IDP, global RCMaaS MySQL.
• Coordinated unit-of-work: a per-store transaction across SQL Server + MySQL with all-or-nothing global rollback and per-database failure identification — practical atomicity without 2PC.
• Pre-flight consistency check: reconciles the max support-user ID across 11+ shards in parallel, alerting on drift before mutating (reconcile-before-write).
• Scope-aware revocation (global vs RCMaaS-linked = least privilege); Serilog audit trail.

Safely replays failed messages from an Azure Service Bus dead-letter queue back onto the topic, in controlled batches.

• PeekLock + complete-only-after-successful-send → zero message loss (at-least-once).
• Preserves MessageId / CorrelationId / Subject / ApplicationProperties → idempotency-safe replay.
• Backpressure-aware: polls the subscription's active-message count and waits for it to drain between batches so replay never re-overwhelms consumers.

Turns a multi-step clearing-house switchover into a single observable, self-notifying job.

• Read accounts from a Google Sheet → switch routing config to DXC in MySQL → push processing queue message → monitor queue depth until drained (not a fixed sleep) → switch back to CHC → Google Chat success/failure notifications.
• Ports-and-adapters (swappable, testable integrations); the revert doubles as the compensating action on failure.