Software & Data Engineer grounding messy, real-world systems in evidence.
This resume cites its sources. to inspect the evidence behind every claim.
01 Experience
Digital Projects Intern to Engineer (Part-Time)
Built the agentic backend for the company's AI energy-analytics chatbot: a LangGraph pipeline with a fetch-then-analyze flow on the Claude SDK and Model Context Protocol tools over FastAPI, so it handled messy real-world queries instead of breaking on anything not explicitly coded for. Attested
Shipped an anti-hallucination suite to production: provenance tracking (EvidencePack/StepResult), a numerical-claim validator that flagged ungrounded numbers, a 5-tier reliability scale, and a citations parser/resolver, so analytics answers could be traced back to the source sensor data. Attested
Built a staged extract/transform/load toolkit for correcting production ClickHouse sensor data: every edit happens in parquet, a single load module is the only ClickHouse write surface, and delete-plus-reaggregation is emitted as reviewable SQL the toolkit never executes. Ingest streams one sensor at a time behind a server-memory gate, after an earlier batch-loading approach overloaded the server. Artifact
Solved correction-versus-dedup correctness on a ReplacingMergeTree table by stamping ingested_at at insert so a correction deterministically wins FINAL, with time-spread read-back verification gating every run before it reports success. Attested
Prototyped and tested a statistical anomaly detector (GESD, MAD/robust-z, IQR fences, per-sensor delta floor, spike-cap safeguard) covering degenerate cases, kept as an option for active threshold filtering, alongside confirmation-phrase and pre-delete backup guards on destructive data paths. Attested
Authored the EECA compliance-report engine (chat-driven editing, browser-editable output, soft delete with auto-expiry) and built its frontend in SvelteKit and Svelte 5: streaming responses, inline artifact cards, and ECharts visualizations. Attested
02 Projects
Tuition-Centre Management Platform · Go · vanilla JS · PostgreSQL
Solo-built and deployed to production: role-based dashboards, billing/payroll, attendance, analytics, notifications, and JWT auth with email verification.
Agentic learning sandbox · Python
A personal sandbox for agent patterns: a ReAct loop with tool execution, retries, and streaming across both Ollama and Gemini. Where I prototype the ideas the CobiNeural work productionises.
Interactive Data Stories · Vega-Lite · JavaScript
Two single-page visual narratives built from raw public CSVs: a choropleth world map, parallel-coordinates crime trends, a state homicide stream graph, radial prisoner-index charts, and case heatmaps.
Co-op terminal game · Java · OOP
A terminal-controlled co-op game (Lethal Company clone) built with three teammates. Wrote the enemy AI behaviour and loot mechanics.
Sequence models · PyTorch
Trained RNN, LSTM, and Transformer models with attention for text generation and sequence classification, then compared how each performed.
03 Skills
Languages
AI / ML
Backend
Frontend
04 Education
05 Leadership