AI & Machine Learning Engineering · OCR, retrieval agents and evaluation
At Redica Systems, as Data Engineer II (MLE), I own a document-intelligence project spanning OCR, structured extraction, retrieval and service integration. I check OCR output page by page, retry when needed and flag unreliable content for review. My extraction workflows validate required fields.
I also build retrieval agents that use GraphRAG and text-to-SQL to query graph and structured data. I expose retrieval through MCP tools and use agent memory to retain context across interactions. I work with Product and distributed engineering teams to deliver these features through FastAPI services and AWS.
Previously, as Software Engineer – ML at Zolvit (formerly Vakilsearch), 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.
Open to senior engineering roles across applied AI, generative AI and machine learning. Bengaluru is my first choice, followed by Chennai. Available immediately.
I build independent testing and evidence tools for agent controls. Start with ARCI to see an experiment, a regression decision, failure diagnosis and offline replay in one runnable fixture.
Watch two 70-second fixture demos · Engineering evidence sheet
The recordings show actual local CLI output, with time compression and explanatory annotations. Their runtime commits, asset hashes and limitations are included.
Runs predeclared baseline/candidate experiments under tool faults, applies non-inferiority gates and replays recorded failures.
Tests a frozen categorical action policy against labelled cases, bounds accepted-action risk and coverage, and replays the evidence.
Turns explicit code claims into executable checks and revalidates them after changes.
Checks declared PR scope against base-commit policy and compiles policy into Claude Code permissions and hooks. Research preview.
Evaluates supplied gates, claims and evidence against acceptance policy. Standalone feature development has ended; the kernel and research record remain available.
Requires macOS/Linux and uv. First setup may download Python and packages; the seeded fixture makes no model calls.
uvx --python 3.12 --from 'agent-reliability-ci==0.8.0' arci demo --workers 4The demo removes a retry, checks the regression, reduces the injected fault set, replays the failure and checks a repair. It exits 0 when the expected outcomes hold, including the intended BLOCK for the broken candidate. Read the fixture and its limits.
The projects' fixture and benchmark results have defined scopes; they do not establish production safety or customer adoption. Each repository documents its trust boundaries and limitations. I use AI coding assistants as pair-programmers.