Agentic AI
- LangGraph
- LangChain
- MCP
- Multi-agent systems
- Tool calling
- Agent memory
- Human-in-the-loop
- Guardrails
- CrewAI
Open to AI engineering roles
Tiju Raju Lukose · AI Engineer
2+ years at Hexaware Technologies shipping FastAPI services, retrieval systems and LLM agents for financial services clients. I work on what makes AI reliable: evaluation, guardrails, observability and cost.
Stack I ship with
End-to-end AI systems, each with an evaluation suite, tests and CI.
Flagship Health insurance
Agentic copilot for a hospital insurance desk. Turns admission and discharge records into insurer-ready cashless claims, predicts deductions before submission and answers insurer queries with cited evidence. Every action on an insurer portal goes through human approval.
Flagship Pharma compliance
Audit-trail analytics and investigation agent for quality-control labs. Flags data integrity red flags across chromatography and LIMS systems, drafts deviation reports with cited evidence and prepares sites for FDA inspections. Runs on premises with open models.
Flagship Privacy
API, SDK and MCP server that finds and redacts Indian personal data in 10+ languages, code-mixed text and ID-card images before it reaches an LLM. An agent handles DPDP access, correction and erasure requests across company systems.
Ask a Postgres database questions in plain English. The agent finds the right schema, writes and safety-checks SQL, runs it read-only, fixes its own errors and returns a table, chart and explanation.
Benchmarked on BIRD Mini-Dev with ablations
OpenAI-compatible proxy in front of many model providers with per-key budgets, rate limits, cost-aware routing, retries, circuit breakers, fallbacks, and exact plus semantic caching.
Load-tested with k6, chaos-tested for provider outages
Indian GST invoices from scans, photos and PDFs to validated JSON with GSTIN checksum and tax arithmetic checks. A LoRA fine-tuned open vision-language model, compared with frontier APIs on accuracy, cost and speed.
Field-level F1, cost per 1k docs, p95 latency
A GitHub App. Label an issue and a sandboxed agent reproduces the bug with a failing test, patches the code, runs the suite and opens a draft PR with evidence.
Evaluated on SWE-bench Verified Mini
Long-term memory any agent can plug into over REST or MCP. Extracts typed memories, tracks when facts change, retrieves with hybrid, recency and entity search, and fully erases a user's data on request.
Evaluated on LongMemEval
Phone and web voice agent that books, moves and cancels clinic appointments in Hindi, English, Marathi and code-mixed speech, with barge-in, a per-turn latency budget and safety escalation.
160 simulated-caller runs across 4 language modes
Also built
Navi Mumbai, India
May 2024 to present
Recognition
Education and certifications
Hiring for AI engineering roles? I'm open to Pune, hybrid and remote positions.