Available — Summer / Fall 2025

I build AI-native
systems infrastructure.

Columbia CS+Mathematics. Five systems across agent orchestration, retrieval pipelines, quantitative engines, and developer tooling — from typed execution graphs to C++ distributed caches. Every claim evidence-backed.

Strong Signal

Abel — typed DAG executor for multi-step LLM pipelines. Systems architecture + AI-native design thinking.

Shipped

Theib — live developer tool running in Claude Code. Real usage, zero setup.

Proof-Backed

Every metric has a source. Every claim has an artifact. Evidence-first, no unattributed numbers.

Executive Brief◈ verified
SchoolColumbia University
DegreeCS + Mathematics
GPA3.7
LocationNew York, NY
StatusOpen to Work
5Systems Built
1Shipped Tool
15+Technologies
10+Backed Claims
Open to
AI infrastructureData engineeringQuant systemsDeveloper toolingSystems programming
Start Here

Five systems. Fifteen minutes.

Columbia CS+Mathematics student building AI-native systems from the stack up. Here's what each project demonstrates and why the order matters.

01
AbelAI Agent Infrastructure
Flagship

The centerpiece. A typed DAG executor for multi-step LLM pipelines — explicit control flow, Pydantic schema validation at every node boundary, Redis-backed distributed state, and a unified adapter across OpenAI, Anthropic, and Ollama. Shows: systems architecture, AI-native design thinking, production-grade execution infrastructure.

Read case study →
02
VidCrawlRetrieval Infrastructure
Data Engineering

Video → semantically queryable knowledge, end-to-end. Five-stage async pipeline: transcript extraction, semantic chunking with cosine-boundary detection, OpenAI batch embedding, Pinecone indexing, FastAPI retrieval. 200k+ chunks at prototype scale. Shows: pipeline architecture, retrieval system design, intelligent infrastructure thinking.

Read case study →
03
KalsherOptions Backtesting
Quantitative Systems

Modular options backtesting with honest execution assumptions. Four independent engines — signal, sizing, execution, attribution — each with a typed protocol interface. VIX regime tagging, realistic fill modeling, post-hoc attribution. Shows: mathematical rigor, modular system design, quantitative engineering discipline.

Read case study →
04
TheibDeveloper Tooling
Shipped

A live developer tool — not a prototype. Runs as a Claude Code slash command with zero setup, searching GitHub, Semantic Scholar, and the web in parallel, scoring 20+ sources by license safety, and producing a structured implementation dossier in under 2 minutes. Shows: shipped product, AI tool design, developer workflow thinking.

Read case study →
05
NoDiffCode Intelligence
Research

Research into a class of bugs invisible to test suites: intent-layer regressions. Tree-sitter AST diffing + two-stage LLM reasoning (intent extraction before regression check) catches behavioral changes that look clean in the diff. Shows: research-oriented thinking, novel problem framing, code intelligence architecture.

Read case study →

By the Numbers

21
Systems Built
Across AI, data, and quant domains
10
AI / Data Projects
Agents, pipelines, embeddings, quant
21
Architecture Diagrams
System design documented
19
Evidence-Backed Claims
Every claim has a source
38
Technologies
Across the full stack
18
Skills Documented
With linked proof
14
Verified Evidence Items
Diagrams, benchmarks, writeups
20
Shipped or Prototyped
Not just ideas — built systems
More Work

Additional Projects

Experiments, utilities, and smaller systems — each one a deliberate exercise in a specific design problem or technical constraint.

Proof Layer

Every claim in this portfolio is linked to an artifact. Each item answers: what does this prove?

Kind
Status
Project
15 of 15 evidence items
Diagram● Preparing to publish✓
Agent Execution Graph

Typed DAG showing how agent nodes declare inputs/outputs and pass structured state through a pipeline, with visible fallback paths.

What this proves

That agent state is explicit, inspectable, and not buried in prompt strings.

Preparing to publish

This artifact exists locally and will be linked here when published.

Backs 1 claim
"Abel's typed DAG makes agent state inspectable at every step, eliminating the ne…"
Source Code● Preparing to publish✓
Core Engine Source

Python implementation of the graph executor: topological walker, Pydantic schema validation, retry/fallback logic, and Redis state backend.

What this proves

That the typed graph executor is a real, working implementation — not a concept.

Preparing to publish

This artifact exists locally and will be linked here when published.

Backs 2 claims
"Abel's typed DAG makes agent state inspectable at every step, eliminating the ne…"
Write-up● Preparing to publish✓
Adapter Interface Design Doc

Design rationale for the unified LLM adapter protocol: `async def complete(messages, schema) -> schema`. Provider differences are fully encapsulated.

What this proves

That the provider-agnostic adapter design was reasoned and documented, not accidental.

Preparing to publish

This artifact exists locally and will be linked here when published.

Backs 1 claim
"A unified adapter interface makes switching between OpenAI, Anthropic, and local…"
Benchmark● Local demo✓
Pipeline Latency Benchmark

End-to-end timing for transcript fetch, chunking, embedding (batched), and Pinecone upsert on a 60-minute video. Local benchmark on prototype infrastructure.

What this proves

That the full pipeline — fetch, transcribe, chunk, embed, upsert — runs in usable time.

Local demo available

Runs on my machine. Live demo or screenshare available on request.

Backs 2 claims
"The full pipeline — transcript fetch, semantic chunking, embedding, and vector u…"
Write-up● Preparing to publish✓
Chunking Strategy Comparison

A/B comparison of naive sentence splitting vs. semantic-overlap chunking on MRR@5. Prototype evaluation on 20 hand-annotated queries across 5 videos.

What this proves

That semantic chunking is a deliberate choice with measured retrieval impact.

Preparing to publish

This artifact exists locally and will be linked here when published.

Backs 1 claim
"Semantic chunking with cosine-similarity boundary detection outperforms fixed-si…"
Diagram● Preparing to publish✓
Four-Module Architecture Diagram

System diagram showing the clean boundary between SignalEngine, SizingEngine, ExecutionEngine, and AttributionEngine — each with a defined protocol interface.

What this proves

That the four-module separation is an actual architectural boundary, not just a claim.

Preparing to publish

This artifact exists locally and will be linked here when published.

Backs 1 claim
"Kalsher's strict module boundary (SignalEngine → SizingEngine → ExecutionEngine …"
Benchmark● Local demo✓
SPX 0DTE Backtest Results

Local benchmark: comparison of credit spread P&L with perfect fills vs. realistic spread modeling, segmented by VIX regime. Illustrates edge degradation and regime sensitivity.

What this proves

That execution timing assumptions materially change strategy evaluation results.

Local demo available

Runs on my machine. Live demo or screenshare available on request.

Backs 2 claims
"Backtests on SPX 0DTE credit spreads show measurable edge degradation when execu…"
Live Demo● Preparing to publish✓
Live Claude Code Skill Demo

Theib running as the `/theib` slash command inside Claude Code, generating a real dossier for a feature request. Demonstrates end-to-end pipeline from query to structured output.

What this proves

That Theib runs as a real in-context tool, not a described-but-unbuilt system.

Preparing to publish

This artifact exists locally and will be linked here when published.

Backs 2 claims
"Theib produces an actionable implementation dossier in under 2 minutes, replacin…"
Write-up● Preparing to publish✓
License Safety Scoring Rubric

Documented scoring rubric for license classification: permissive (MIT, Apache 2, BSD), weak-copyleft (LGPL, MPL), strong-copyleft (GPL, AGPL), and unknown. Sources with score < 0.5 are excluded from the dossier.

What this proves

That license classification is rule-based and reproducible, not a vibe check.

Preparing to publish

This artifact exists locally and will be linked here when published.

Backs 1 claim
"Automated license scoring filters GPL-contaminated and restrictively licensed so…"
Write-up● Preparing to publish
Regression Detection Case Study

Prototype analysis of 10 real-world regressions: changes that passed all tests and looked clean in code review, but violated behavioral invariants detectable by two-stage LLM reasoning. Planned formal evaluation pending.

What this proves

That semantic regressions exist in the wild and are reproducibly missed by tests and diffs.

Preparing to publish

This artifact exists locally and will be linked here when published.

Backs 1 claim
"Two-stage LLM reasoning (intent extraction before regression check) catches beha…"
Diagram● Preparing to publish✓
AST Diff Pipeline Architecture

Architecture diagram of the three-layer pipeline: structural layer (Tree-sitter AST diff), intent layer (Claude intent extraction), regression layer (Claude invariant check with chain-of-thought).

What this proves

That the AST pipeline is a real structural system, not a wrapper over text diffs.

Preparing to publish

This artifact exists locally and will be linked here when published.

Backs 2 claims
"AST-level diffing eliminates formatting noise from line-level diffs, producing a…"
Write-up● Preparing to publish✓
Evaluation Loop Architecture

Design document for the evaluator-optimizer loop: rubric design, prompt versioning schema, performance attribution methodology.

Preparing to publish

This artifact exists locally and will be linked here when published.

Backs 1 claim
"The evaluator-optimizer loop measurably improves task performance over successiv…"
Benchmark● Local demo✓
Improvement Curves

Performance curves across iterations on a code generation task. Local benchmark — shows measurable gain from baseline to optimized prompt versions.

Local demo available

Runs on my machine. Live demo or screenshare available on request.

Backs 1 claim
"Structured rubric-based evaluation produces reproducible scores, making improvem…"
Source Code● On request✓
C++ Implementation

Full C++ source: TCP server, RESP protocol parser, shard-locked concurrent hashmap, O(1) LRU doubly-linked list, TTL background worker, write-ahead log.

Available on request

Source shared with companies on request. Available for technical interviews and screenshares.

Backs 2 claims
"Thread-safe concurrent access is achieved without a global lock: 256-shard mutex…"
Benchmark● Local demo✓
Throughput Benchmark

Local benchmark: reads/sec and writes/sec under concurrent load (8 threads, 100k operations). Compared against naive single-threaded implementation.

Local demo available

Runs on my machine. Live demo or screenshare available on request.

Capability Map

Core skills with linked project evidence — ranked by relevance for your hiring context.

AI & ML
AI Agent SystemsExpert

Typed execution graphs, multi-step reasoning, fallback logic, evaluation loops

LLM OrchestrationExpert

Multi-provider adapters, structured output, prompt versioning, evaluation

Embeddings & Vector SearchProficient

OpenAI embeddings, Pinecone, semantic chunking, retrieval

NLPProficient

Semantic chunking, Whisper, text classification, intent extraction

Knowledge SystemsProficient

Knowledge graphs, retrieval pipelines, structured synthesis

Vector SearchProficient

Pinecone, namespace management, filtered retrieval

Systems
Code IntelligenceFamiliar

AST parsing, structural diffing, semantic analysis

C / C++Proficient

Concurrency, TCP networking, POSIX threads, memory management

Distributed SystemsProficient

Concurrency, shard-locked data structures, replication, consistency

Systems ProgrammingProficient

Event loops, thread pools, wire protocols, write-ahead logging

Backend
API DesignExpert

REST, FastAPI, type-safe interfaces, protocol design

PythonExpert

Async, Pydantic, dataclasses, scientific stack

Data
Data PipelinesExpert

Async ETL, batch processing, fault-tolerant orchestration

Quant
Quantitative FinanceProficient

Options strategies, P&L attribution, risk modeling

BacktestingExpert

Event-driven simulation, realistic execution modeling, regime analysis

Options & DerivativesProficient

Greeks, credit spreads, iron condors, 0DTE strategies

Statistical AnalysisProficient

Sharpe ratio, drawdown analysis, regime classification

Frontend
TypeScriptExpert

Strict mode, type-driven design, generics

Daniel Boehme
Let's talk
Daniel Boehme

AI-native systems engineer. Columbia CS+Math. Open to work.

Actively looking for roles starting summer/fall 2025. AI infrastructure, data engineering, quantitative systems, and developer tooling. Reach out directly — I respond quickly.

AI infrastructure & agent systemsData engineering & knowledge extractionQuantitative systems & algorithmic researchDeveloper tooling & code intelligenceSystems programming & infrastructure