Projects

Projects

Built are systems I built myself: what I did, how they work and how success was measured. Led are projects built by teams I led as lead engineer, credited to the team. Client names are withheld.

BuiltLLM systemsInsurance
01

Comparing five insurers’ policies, clause by clause

A retrieval-free extract-then-arbitrate pipeline over long policy documents.

68% → 91%Accuracy against an expert’s answers, from off-the-shelf RAG to the final pipeline
Read the case study
BuiltLLM systemsB2B SaaS
02

Insight generation for a customer-intelligence platform

A judge validated against hand labels, 80–90% wasted reprocessing removed, incidents traced to root cause.

90%+Judge agreement with hand-labeled data on the two checks validated
Read the case study
BuiltPlatformArchitecture
03

Event-driven platform for algorithmic crypto trading

An architecture where one algorithm runs unmodified on live data and in historical replay.

Live = replayDesign goal: no separate backtest logic per algorithm
Read the case study
BuiltDeep learningNLP
04

Reproducing a landmark question-answering model

BiDAF implemented from the paper, with attention flowing between passage and question.

83.5 EMSQuAD v1.1, within about a point of BERT’s reference
Read the case study
BuiltLLM systemsCreator economy
05

Persona architecture for creator AI clones

One reusable prompt template plus persistent memory keeps every creator’s clone consistent.

Read the case study
LedLLM systemsDocument extraction
06

Teaching a small model to say “I don’t know”

SpecSense: fine-tuning a sub-1B model to extract product data from messy documents, cutting invented answers from 38% to 0.03%.

Read the write-up
LedLLM systemsExplainer video
07

Plan, don’t generate: an AI that directs a video but never builds it

Canvas Motion: a topic in, a narrated animated explainer out, with the AI choosing from about 59 hand-built components.

Read the write-up
LedPlatformVoice AI
08

Vāṇī: a voice AI engine where every part swaps

A real-time voice engine with natural interruption that runs in the cloud or fully offline from one codebase.

Read the write-up