~/hassan-pasha_[ Book a Call ]
hassan@prod:~$ whoami — Principal AI Engineer & Head of AIChicago / LLM · Vision · Inference

AI that
survives
production_

Principal-level AI engineer who sets strategy and ships the models — LLM fine-tuning and serving, computer vision, and inference on GPU fleets. Founded Sequence ($5M exit). Building agents since Theory of Mind research in 2015 — results measured in production systems, not demos.

[ Book a strategy session ][ View the work ][ Resume PDF ]
$5M
Exit — Sequence CV & agent platform
8.3M
Full-text scholarly graph nodes
27B
DBSA reasoning model trained
11 yrs
Building AI — since 2015

$ open ./case-studies # 01

Don't take my word for it.
Walk through the work.

Nine systems, newest first — each one a real deployment with the problem, what I built, and what it returned. Pick one.

DBSA-27B, an 8.3M-node scholarly graph, and long-context reasoning infrastructure

27B
DBSA reasoning model
8.3M
Full-text graph nodes
530.7M
Graph edges
## The problem

Frontier models are bottlenecked by training data — especially long-context. Models advertise huge context windows, but the data that teaches genuine long-range reasoning is scarce, expensive, and hard to validate at scale.

## What I built
  • —DBSA-27B — continued pretraining and instruction tuning of a 27B reasoning model on long-context reasoning-over-code corpora
  • —Full-text scholarly knowledge graph on the OpenAlex spine: ~8.3M full-text nodes and 530.7M edges from arXiv, PMC, bioRxiv, and medRxiv
  • —Hybrid dense/sparse retrieval and reranking for graph-aware RAG over real scientific text
  • —A deep-research engine that synthesizes multi-source evidence with per-source quality scoring and citations
“The next capability jump in AI is not more parameters — it is better long-context data.”
Classified — full build file
Week-by-week implementation, architecture, and the complete returns breakdown. Passcode holders only.
[ Attempt access ]
LoRATensorRT-LLMSGLangRAGOpenAlexTraining Data
Visit site [ I have a problem like this ]
$ ls ./more-work
LuxUp
P2P luxury rental marketplace
2020
AI Domain Registrar
First AI-powered registrar
2020
Elderly-Care AI
Built on Theory of Mind research
2020

$ grep "your problem" ./solutions # 02

Which one is costing you money?

I don't sell technology. I remove specific, expensive problems — and every one below is backed by a deployment you just saw.

Your models are too slow or too expensive to serve
TensorRT-LLM / SGLang / vLLM inference — LoRA adapters, quantization, continuous batching on H100 fleets
→ Aldea: production GPU serving + 512-GPU diligence
Your RAG can’t ground answers in real source text
Full-text scholarly graphs and hybrid dense/sparse retrieval over parsed papers
→ SubQ: 8.3M nodes · 530.7M edges
Voice agents feel laggy or locked to a vendor API
LiveKit streaming ASR → LLM → TTS on owned infrastructure, sub-second turns
→ Aldea: real-time voice AI in production
You can’t see threats until it’s too late
Real-time computer vision that detects, scores, and classifies events as they happen
→ Sequence: $5M exit · 87% faster response
Your cloud bill grows faster than revenue
Forecasting and automated remediation that eliminates waste before the invoice lands
→ CVS: $10M+ saved per year
Your AI pilots never reach production
End-to-end engineering — models, pipelines, backend, and infrastructure that ship
→ 11 years, every system in production
[ My problem isn't listed ]30 minutes. If AI won't fix it, I'll tell you.

$ grep -r "results" ./clients # 03

In their words

“Hassan's AI integration reduced our contract processing time by 72% and increased our team's case capacity by 40%. The ROI was clear within 60 days.”
David Miller
Managing Partner, Fitter Law
72% faster document processing

$ cat how-it-works.md # 04

What happens after you book

No mystery. Four stages, each with a defined cost, a defined deliverable, and a decision point before the next one starts — commitment on both sides.

01./scope

Scope call

30 min

You bring the business problem. I tell you honestly whether AI solves it and what class of investment it takes. If it’s a fit, we price the discovery phase on this call.

02./discover

Paid discovery

~30 days

A fixed-fee, fixed-window engagement, priced to the size of the system being scoped. Architecture, data audit, ROI model, and a phased delivery roadmap. Serious projects start here — discovery is never free.

03./prove

Proof of concept

weeks 1–8 of build

A working demonstration on your real data before full production funding. You see the system operate — not slides about how it will.

04./deliver

Phased delivery

quarterly milestones

Full platforms take a year or more — so delivery is staged. Every quarter ships a meaningful, working result, and every phase is funded by the results of the last.

$ ls ./engagement-models # 05

Three ways to work together

Every engagement starts with paid discovery and moves in funded stages — you commit to one phase at a time, and every phase ends with something working. Pricing follows the scale of the system, and we get to a number on the first call.

4.1
~30 days

Discovery

Paid scoping, fixed fee

A fixed-fee discovery engagement, priced to the scale of the system on the table. You leave with an architecture, an ROI model, and a phased roadmap you can fund with confidence — whoever builds it.

  • —System architecture & technical blueprint
  • —Data and infrastructure audit
  • —ROI model with measurable targets
  • —Phased delivery roadmap with milestone pricing
  • —Go / no-go recommendation — honestly given
For — Serious organizations committing to AI
4.2
3–8 weeks post-discovery

Proof of Concept

The system, demonstrated on your data

The core of the system built and demonstrated against your real data and workflows — de-risking the full build before production investment. Even a focused demo takes weeks to do honestly.

  • —Working PoC of the highest-risk component
  • —Evaluation against real data, not toy sets
  • —Demo your leadership can see running
  • —Refined production plan and pricing
For — Proving the thesis before scaling investment
4.3
12+ months, staged

Build & Partner

Phased delivery, quarterly results

Full production platforms take a year or more, built in funded phases. Every quarter delivers a meaningful, working capability — you re-commit each phase based on shipped results, not promises.

  • —Quarterly milestones, each independently valuable
  • —Models, pipelines, backend, and infrastructure
  • —Team training and documentation per phase
  • —Ongoing operation, monitoring, and optimization
  • —Priority access and strategy sessions
For — Enterprises building AI as core strategy

Not sure which shape fits? That's what the first call is for — we'll scope the problem and I'll tell you honestly what it takes, including if the answer is “you don't need me for this.”

Scope your project

$ ls -la ./vault # 06 — permission denied

The vault

Every engagement produces a full build file — week-by-week implementation, architecture decisions, and the complete returns breakdown. Clients get a passcode with their contract. Everyone else gets this wall.

subq.build
AI Research
Access: restricted[ attempt ]
aldea.build
Voice AI
Access: restricted[ attempt ]
wealthiq.build
Fintech
Access: restricted[ attempt ]
sequence.build
Security
Access: restricted[ attempt ]
rfp-automator.build
Gov Contracting
Access: restricted[ attempt ]
cvs-cloud.build
Fortune 500
Access: restricted[ attempt ]
fitter-law.build
Legal
Access: restricted[ attempt ]
servusx.build
Service Ops
Access: restricted[ attempt ]
thexcrm.build
Sales
Access: restricted[ attempt ]
covid-epp.build
Public Health
Access: restricted[ attempt ]
cpdai.build
Public Safety
Access: restricted[ attempt ]
$ echo "passcodes ship with signed contracts"[ Get clearance → book a scope call ]

$ git log --since=2015 # 07

Eleven years of shipped AI

From Theory of Mind research to a $5M exit, a 27B reasoning model, and an 8.3M-node scholarly graph — every engagement went to production with measurable results. No proofs-of-concept that died in a slide deck.

The through line: agents before LLMs, LLMs before ChatGPT, and production inference infrastructure before it was a job title.

[ Download full resume → ]
2015–17

Doctoral Research — Theory of Mind & Machine Cognition

UIUC research into computational models of belief, intent, and reasoning about other minds — the capabilities now central to modern LLMs. Independent work on AI agents built for a 1:1 relationship with their human, before LLMs existed. B.S. Computer Science, University of Illinois Chicago (2017).

2017–18

Lead Engineer — Advance Technology Services

Custom NLP over millions of scraped RFP PDFs — tokenization, bag-of-words, and ETL that helped close millions in won RFPs and cut response time in half. Pre-GPT language automation for government contracting.

2018–21

Lead AI Developer → CTO → Advisor — Evolution Business

Built and trained computer-vision and deep-learning models (TensorFlow) for healthcare SaaS; owned data curation, model accuracy, and distributed training/serving. Stood up ETL batch pipelines (Airflow, Databricks) and predictive models that increased lead generation. Computer vision and NLP pipelines for City of Chicago urban infrastructure.

2020

Founder — three products shipped

Elderly-care AI built on Theory of Mind research, a COVID tracking system used by the Illinois Department of Health, and the first AI-powered domain registrar. LuxUp: peer-to-peer luxury rental marketplace with QR-verified pick-up workflows.

2020–21

Lead ML Engineer — NOCD

ML-backed matching algorithms (Python, Flask, React) and HIPAA-aware data architecture on PostgreSQL, DynamoDB, and Redis for a revenue-critical patient application.

2022–23

Principal Solution Architect, AI / Cloud — CVS Health

Architected secure, serverless AI systems on AWS (Lambda, Step Functions, API Gateway, SageMaker). Built ETL and feature pipelines feeding ML models; automated security measures that cut non-compliant service creation by 70%. Led a team of 7; cloud cost forecasting and remediation that saved $10M+ annually as the team scaled from 2 to 30. Hands-on through 2023; the company engagement remains active.

2023–25

Sequence — Computer Vision & AI-Agent Platform · $5M Exit

Founded Sequence: real-time threat- and activity-detection — pose estimation, keypoint extraction, skeleton normalization, and temporal transformers (MMAction2). Multi-stage AI-agent workflows over FastAPI with Dockerized serving on AWS. Acquired in a $5M exit; platform live across finance and enterprise networks with 87% faster breach response.

2024–25

ServusX, TheXCRM

An all-in-one ops suite for service businesses and an AI-driven CRM — built, shipped, and operating.

2025–

Founding Head of AI → Principal AI Engineer — Aldea → SubQ

Founding team / A-class shareholder. Led AI research and long-context training-data engineering: DBSA-27B domain-adapted reasoning model, LoRA adapter serving, and TensorRT-LLM / SGLang / vLLM inference on H100 fleets. Built an 8.3M-node full-text scholarly graph (530.7M edges) on OpenAlex for graph-aware RAG. LiveKit real-time voice AI agents. Evaluated GPU fleets (H100–B300) and helped close a 512-GPU cluster deal. Aldea acquired by SubQ; continuing principal AI work plus WealthIQ.

$ tail ./writing # 08

Notes from the field

01
Enterprise AI Implementation Strategy: The 2026 Playbook
Enterprise AI · Strategy
Jul 2026 · 4 min
02
Why Most AI Pilots Never Reach Production — and How to Ship Yours
Enterprise AI · Implementation
Jun 2026 · 5 min
03
How to Measure Enterprise AI ROI: A Framework With Real Numbers
ROI & Metrics · Enterprise AI
May 2026 · 4 min
All articles

$ ./contact --book-session # 09

Have a system that needs to exist? Let's scope it.

A 30-minute scope call. You bring the business problem; I'll bring an honest read on whether AI solves it and what class of investment it takes. If it's a fit, we scope the discovery engagement on the spot.

Book the sessionOr reach out on LinkedIn