Ingyu Koh

My solution recovered license plate numbers from high-speed video footage where motion blur, low resolution, and poor lighting defeated standard OCR. The system selected and aligned the usable frames from each plate, combined them to achieve a higher effective resolution, and applied character recognition tuned to the plate's typography rather than general text. The system reported the result as a confidence assessment rather than a single answer: which characters could be established, which remained genuinely ambiguous, and why. That distinction matters when output may be relied on as evidence, and it is where most recognition work is weakest. The solution delivered reconstructed imagery together with a written analysis that the client could act on. The system rated 5.0 on completion.

I audited quantitative research whose conclusions rested on statistical arguments that the authors could not independently verify. Rather than accepting the stated results, I re-derived the mathematics from first principles, reproduced the calculations, and tested whether the statistical methodology actually supported the claims being drawn from it.

The deliverable separated three things most reviews conflate: what the audit could confirm (mathematical consistency and correct calculation), what it could not (whether the model corresponds to reality), and where the methodology itself was the weak link. Explicitly stating that boundary is what makes an audit useful rather than merely reassuring. I delivered a written report that the client used to revise the work.

I built the research infrastructure that turns market hypotheses into repeatable experiments. Tick and limit-order-book histories are large enough that a naive backtest runs for hours, which quietly limits how many ideas ever get tested. Moving simulation and feature computation onto the GPU cut that loop from hours to minutes and made large-scale model training practical on the same data.

The platform enforced the discipline that matters more than speed: strict point-in-time data access so no future information can leak into a signal, realistic fill and latency assumptions, and rolling-origin evaluation instead of a single train/test split. Signals studied included spread behavior, quote and depth imbalance, and short-horizon price formation.

The result was a shorter path from question to evidence, and fewer results that looked good only because the test was wrong.

Contributed to DIVEROID’s iOS application development for an iPhone used underwater inside a protective enclosure. My work included Swift development, user-interface development, and bug fixing for this specialized underwater smartphone application.

I developed a portfolio recruiting assistant with an MCP server, LangChain MCP adapters, and LangGraph for multi-turn tool queries. I added bounded tool loops, prompt-injection and PII guards, FastAPI and Docker serving, and human decision support. On October 6, 2026, I reproduced 27 passing tests and ran a synthetic 13-conversation, 30-turn offline evaluation, achieving 100% conversation success with guards compared with 61.5% without them. Pattern-based guards are not exhaustive. The reported results use a deterministic offline planner; I also implemented an optional Claude adapter, but no Claude runtime result is claimed. This is recent independent engineering evidence, not prior client or enterprise production usage. Source, tool trajectories, raw results, and setup: github.com/ingyukoh/mcp-recruiting-agent.

Built a reproducible LangGraph and FastAPI portfolio system for numeric questions over SEC 10-K filings. I implemented filing ingestion, BM25 retrieval, evidence extraction, provenance checks, abstention, the evaluation harness, tests, Docker, and CI. On a rebuilt 64-question offline benchmark using real filing text, task success was 57.8%, compared with 12.5% for the BM25 baseline. The earlier 100% versus 65.6% benchmark was retired because a simple fact lookup could saturate it. This benchmark uses deterministic control flow rather than an LLM, and the system was not deployed for a client.

Ingyu built an independent Python portfolio system that routes bounded tools through LangGraph, retrieves cited evidence, checks prompt-injection patterns, and records per-node traces. The implementation includes FastAPI, Docker, tests, and a GitHub Actions workflow that runs lint, tests, and an offline evaluation on pushes and pull requests. The evaluation compares the agentic route with a non-agentic baseline using a small corpus and a deterministic provider; OpenAI Responses is optional. Source, CI configuration, traces, and reproducible setup are inspectable in the linked repository. This demonstrates tool orchestration, API development, and continuous integration. Enterprise agentic SDLC deployment, consequential-action approval gates, and NeMo/Guardrails AI implementation are not established by this prototype.

In October 2026, Ingyu developed an independent Legal Contract Review Lab with AI assistance using Python, PyTorch, Hugging Face Transformers, and PEFT LoRA. Qwen2.5-0.5B-Instruct was adapted for yes/no classification of agreement-ending paragraphs; quotations and line/character locations are copied from the source. On a separate synthetic challenge, the adapter scored 95% on clauses and 76% including headings, versus 85% and 68% for the frozen base. Source-copy validity was 100%; relevant-quote precision was 62.5%. Model and adapter manifests, pinned revision, and artifact hashes support reproducible loading. Docker/FastAPI on private EC2 sits behind an AWS Lambda HTTPS entry point, with bounded inputs, request limits, health checks, missing-model responses, and restart recovery. Six serial AWS reviews had a median latency of 7.13 seconds and a peak container memory of 3.21 GiB. This demonstrates a basic model lifecycle, evaluation, and serving, not enterprise MLOps or prior law-firm operation. The public examples and custom-text demo need human review. Scheduled stop: 11 October 2026, 14:45 Seoul.

In October 2026, I developed this independent research prototype with AI assistance using Python and scikit-learn. It uses 989,895 public Kalshi Bitcoin-binary trade prints, compares logistic and gradient-boosted hypothetical fill models, and models fill quantity and adverse selection separately. Chronological evaluation covers 45 markets; 11 markets across two days are held out. Boosting achieved inferred-fill Brier 0.2236 versus constant 0.2473. Replay includes partial fills, inventory limits, tick sizes, fees, latency/queue sensitivities, and a frozen delayed-feedback LinUCB comparator. The learned policy lost $24.64 versus $4.44 for the validation-selected fixed policy; fills are tape estimates, not observed own-order fills. The public AWS API provides three editable examples, source download, data provenance, full results, and failure cases. 18 checks passed. Measured HTTPS median was 798.5 ms versus 4.9 ms API computation. This is recent crypto market research, not a profitable production engine or an exchange execution service.

On October 6, 2026, Ingyu developed an independent Python business-decision prototype with AI assistance: frozen OLS, a small SQLite semantic graph, and Qwen2.5-0.5B-Instruct (494,032,768 parameters, no fine-tuning). Sales RMSE on 40 held-out Advertising rows was 1.9761 versus 5.8287 for the training-mean baseline. After an adversarial review exposed a bypassable keyword check, Ingyu replaced it with a closed display policy: 30 invalid probes authored were rejected, and 10 valid controls were accepted. Numeric results and units come from code; raw LLM text remains untrusted diagnostic evidence. Twenty recorded AWS generations are available for inspection; preset examples are cached, and the inference host is stopped when idle. Custom numeric predictions remain available. The implementation includes FastAPI, Docker, bounded requests, health checks, fallback/restart handling, and two read-only MCP tools. On October 6, 21 gateway/output tests passed; these are scoped regression checks. Source, recorded outputs, and evaluation: github.com/ingyukoh/business-decision-context-lab. This demonstrates recent prototype engineering without establishing causal ROI or enterprise-agentic SDLC delivery.

In October 2026, Ingyu implemented an independent approval-gated SDLC lab with Python, LangGraph, and a custom Guardrails AI validator. It inspects a disposable invoice repository, validates an untrusted recorded Bedrock Nova Micro patch proposal, waits for human review, and applies only the approved SHA-256 digest. Controls include allowlisted tools/paths, an arithmetic AST policy, 120-second approvals, and atomic single-use claims in DynamoDB. The AWS Lambda HTTPS demo exposes the diff, sanitized action/decision/reason traces, test outcomes, evaluation results, and source downloads. The authored corpus blocked 24/24 forbidden proposals and accepted 12/12 benign variants; 61 container tests and 23 live HTTP assertions passed. Changed, expired, and replayed approvals are tested; a waiting review survived a new worker and was executed once. Nine serial HTTPS mutations had a median of 258 ms. This is recent prototype evidence: planning replays a recorded model response; the executor handles a single bounded file; and enterprise SSO, general sandboxing, and real Git/CI/deployment adapters remain integration work.