Applied AI & Machine Learning · Agentic systems, retrieval and evaluation
I build AI applications that connect language models to data and tools, and open-source software for testing agent behavior. My work includes retrieval integrations, MCP tools and agent memory.
Open to senior individual contributor roles across applied AI, generative AI and machine learning. Bengaluru first, Chennai second. Available immediately.
Problem: an agent change can behave differently when a tool times out or fails.
What I built: a Python CLI and GitHub Action for comparing agent versions on declared tasks and faults, inspecting differences and replaying recorded failures offline.
Try it: the seeded retry fixture checks a broken candidate and a repair. Requires macOS/Linux and uv. Initial setup downloads packages; the fixture makes no model calls.
uvx --python 3.12 --from 'agent-reliability-ci==1.0.0' arci demo --workers 4The command exits 0 when the expected outcomes hold, including the intended BLOCK for the broken candidate. Results apply to the tested tasks and faults; a recorded difference is not proof of root cause. Demo details and limits.
Problem: accepting more actions can also mean accepting more actions that disagree with the labels.
What I built: a Python toolkit that evaluates a fixed policy on labelled cases, computes bounds on accepted-action risk and coverage, and records evidence for offline replay.
Try it: use a new output directory. The demo compares bad and fixed fixtures and checks their replay; no model or API key is needed.
uvx --python 3.12 --from 'actseal==1.0.1' actseal demo --out ./actseal-demoThe complete demo exits 0 after the expected bad-fixture BLOCK and fixed-fixture PASS. Replay checks consistency with recorded evidence; it does not authenticate responses or establish that labels are true.
Watch the recorded fixture demonstrations · Engineering evidence sheet
Recordings retain their original runtime versions, time compression, annotations and documented limits. They are historical demonstrations and may predate the pinned releases above.
Redica Systems · Data Engineer II (MLE)
At Redica Systems, I contributed to the agentic AI platform by integrating GraphRAG and text-to-SQL into retrieval workflows, exposing retrieval through MCP tools, and implementing memory to retain context across interactions.
I also owned OCR engineering from EC2 and model selection through resource optimization, including page chunking, disk-backed intermediates and concurrency controls. This work included quality checks, controlled OCR retries and review signals for unreliable output. Employer implementation and customer material remain private.
Zolvit (formerly Vakilsearch) · Software Engineer - ML
I built hybrid legal-document search, summarisation and extraction systems, and adapted open-source language models through continued pre-training and parameter-efficient fine-tuning.
Working tools: Python, FastAPI, AWS, PyTorch and Docker.
- Dorian: rechecks explicit code claims after changes. A claim is only as strong as its checker.
- Frontier Scout: checks declared pull-request scope against base policy. Research preview.
- Evalopt Graph: applies acceptance rules to supplied evidence. Standalone feature development has ended.
These projects provide bounded testing and evidence; they do not establish production safety or customer adoption. I use AI coding assistants as pair-programmers.