Instinct-Trained AI Models Deployed Across 6 Industries
Osman Janjua
AI Solutions Architect

Osman Janjua

Building production ML pipelines, fine-tuning domain-specific LLMs, and architecting multi-cloud infrastructure across AWS and Azure. Deep learning field extraction, computer vision inventory systems, 55-service orchestration, and end-to-end automation — from PyTorch model training to containerized deployment.

Chicago, IL · [email protected] · 443.799.7399
21
Production Services
6
Industries
4
Cloud Platforms
8
AI Models Integrated
The Instinct Training Method
01

Capture

We sit with domain experts — founders, architects, CFOs, controllers — and extract the decision patterns they can't articulate. Structured interviews. Live observation. Decision journaling. MBTI-based cognitive profiling maps each stakeholder's decision style, so the model learns how they think, not just what they decide.

02

Train

Fine-tune foundation models on captured instinct data. QLoRA, RLHF, domain-specific tokenization. Training data is stratified by decision archetype — we weight examples by cognitive type so the model doesn't over-index on one thinking style. We don't train on generic web data — we train on the decisions that built companies.

03

Deploy

Production deployment on cloud infrastructure. Real-time decision support integrated into your existing workflows. Outputs are personality-tuned per audience — investor-facing docs match analytical profiles, sales copy speaks to experiential ones. A decision engine with the instinct of your best people.

Deep Dive: The Full Pipeline

Click any stage to see the technical breakdown, why it matters, and how the pieces connect.

01

Decision Pattern Extraction

Why this is the foundation everything else depends on

Every model is only as good as its training signal. Generic AI trains on the internet — broad but shallow. Instinct training starts with the specific decision patterns that made your organization successful. We extract what your best people do instinctively — the judgment calls they can't explain but always get right. Each expert is behaviorally typed during extraction — an analytical CFO's risk calculus is captured differently than a visionary founder's opportunity instinct, preserving the cognitive diversity that drives real organizations.

Founder
Pattern recognition
Architect
System intuition
CFO
Financial instinct
Operator
Execution patterns
↓ Structured interviews · Live observation · Decision journals ↓
Decision Pattern Corpus
100K+ signals codified into training-ready format
100K+
Emails analyzed
WhatsApp
Conversation threads
Voice
Tone & cadence
Deals
Win/loss patterns
16
Decision archetypes
02

Foundation Model Fine-Tuning

Turning raw instinct data into a model that thinks like your team

We don't build models from scratch. We take proven foundation models (Qwen 8B, Claude, open-source) and surgically adapt them using QLoRA — low-rank adapters that encode your organization's instinct without losing the model's general intelligence. Training examples are tagged with cognitive decision archetypes — the model learns that an operator's supply chain call and a strategist's market read are both valid but structurally different. The result: a model that reasons like GPT-4 but decides like your best operator.

Foundation Model
Qwen 8B / Claude / OSS
+
Decision Corpus
Your instinct data
↓ QLoRA adapters · RLHF · Archetype-stratified weighting · Domain tokenization · Eval benchmarks ↓
Instinct Model
Your organization's decision intelligence — encoded as weights
QLoRA
Rank 64 adapters
31%
Eval improvement
GPTQ
4-bit quantized
Local
On-prem inference
03

Production Deployment & API Layer

From trained model to live endpoints your systems can call

A model sitting on a laptop is a demo. A model behind an API with auth, rate limiting, failover, and logging is infrastructure. We deploy across Cloudflare Workers, GCP Cloud Run, and Azure — multi-region, sub-200ms, with monitoring and rollback built in.

Instinct Model
↓ Containerized · Load balanced · Auth + Rate limiting ↓
REST API
Sync queries
WebSocket
Stream output
Serverless
Event-driven
Batch
Document stacks
04

UI, Automation & Document Stack Generation

Embedding intelligence into the tools your team already uses

The model doesn't live in a chat window. It's woven into dashboards, document pipelines, voice agents, and messaging channels. When a work order arrives, the model reads it, matches it, routes it, and drafts the response — in your voice, following your rules. The Personality Playbook shapes every output: investor memos use analytical structure, client proposals lead with relational warmth, and ops runbooks match the detail-first precision operators expect. Document stacks that used to take days generate in seconds.

Instinct Model API
↓ Queries the model on every trigger ↓
Dashboards
Live KPIs
Voice Agents
11 channels
Doc Gen
OMs, invoices, SOWs
↓ Automated outputs ↓
CRM Updates
Email Drafts
Pitch Decks
Compliance Docs
05

Continuous Query & Human-in-the-Loop

The model gets sharper with every decision your team makes

This isn't a one-time deployment. The model is queried continuously — every document it generates, every decision it recommends, every automation it triggers passes through human checkpoints. Approvals, rejections, and edits feed back as training signal, each tagged with the reviewer's cognitive profile so the model refines its understanding of which decision styles accept which framings. The model compounds institutional knowledge over time.

Live Query → Model Response → Human Review
↓ Approved? Ship it. Rejected? Model learns why. Edited? Even better signal. ↓
Approval
+1 confidence
Rejection
Negative signal
Edit
Highest value signal
↺ Feeds back to training — next query is sharper

Experience

AI Solution Architect & Consulting Lead

ELA Asset Management — Multi-Industry
Current
  • Architect AI solutions for Fortune 500 clients across energy, healthcare, real estate, and restaurant sectors — 21 production services deployed across 4 cloud platforms
  • Pioneer instinct-training methodology with MBTI-informed personality playbook — capturing decision patterns from founders, systems architects, and financial leaders into production AI models that scale institutional knowledge with cognitive-type-aware training data and outputs
  • Built ERP-integrated document intelligence pipeline processing invoices across 14,000 vendor accounts with multi-currency GL routing, SAP F3041 template generation, and 5-tier approval workflow
  • Engineered computer vision inventory system with PyTorch — reducing manual errors 40% with real-time alerts across 11 messaging channels
  • Fine-tuned domain-specific LLMs (8 AI models integrated, 5,600+ over 10 years of ML training examples) achieving 31% eval loss improvement across healthcare and insurance verticals

Principal, Strategic Acquisitions & Asset Management

ELA Asset Management — Chicago, IL
Current
  • Lead strategic acquisitions across foreclosure, probate, and distressed asset portfolios with automated CMA and structured deal flow
  • Built AI-driven offering memorandum automation — generating institutional-quality OMs from property data, comps, and financials in minutes instead of weeks
  • Develop capital raising strategies including investor relations, fund structuring, LP/GP reporting, and automated fund administration workflows
  • Engineered DST and 1031 exchange document stack automation — compliance-ready packages with QI coordination, identification notices, and closing docs generated programmatically
  • Manage arbitration processes and dispute resolution in complex multi-party transactions involving liens, title defects, and contested ownership

Director, Strategic Solutions

Energy Sector
Previous
  • Led initiatives supporting $200M+ in annual billing — pricing structure, commodity analysis, and operational decision support
  • Drove smart contract deployment and process modernization across commercial infrastructure
  • Supported power plant acquisitions and aligned financial structure with operational execution

Strategy / Operations Transformation Leader

Family Neurology Practice — Maryland
Previous
  • Supported expansion to top-billing practice in Maryland through technology-first operational redesign
  • Redesigned patient throughput, billing workflows, and reporting — driving multi-year sustained growth
  • Applied analytics-driven scaling across scheduling, staffing, and financial performance

Technical & Leadership

AI & Engineering

PythonPyTorchTensorFlow LLM Fine-TuningComputer Vision Deep LearningCognitive ProfilingMLOps AWSAzure Cloud ArchitectureERP Integration

Strategy & Leadership

AI Strategy Executive Communication Team Leadership Business Development Stakeholder Alignment Mentoring Security & Compliance

Featured Work

Production systems, real results

View Full Portfolio →
Production
Enterprise AI

Document Intelligence Pipeline

End-to-end invoice processing with deep learning field extraction, SAP ERP integration, multi-currency GL routing, and 5-tier approval workflow.

Multivendor
SAPintegrated
F3041templates
PythonDeep LearningSAP
F3041 Ready
Production
Computer Vision

Franchise Inventory System

7 AI agents in one network — automated PO generation, Oracle data entry & sync, threshold alerts, supplier verification, PyTorch barcode scanning, and OCR product detection for a Fortune 500 restaurant group.

7AI agents
40%error reduction
PyTorchOracleMulti-Agent
Production
AI Automation

AI Lead Generation Platform

20,000+ towns mapped for SEO and ad campaigns, voice agent qualification, CRM pipeline automation.

20K+towns
8voice agents
ElevenLabsTwilio

Energy Command Center

GIS-powered lead scoring across commercial properties with automated outreach and building electrification strategy.

10Mproperties

Instinct Model Training

Domain-specific LLMs trained on founder and operator decision patterns with archetype-stratified training data. 8 AI models across 6 industries — highest personality-playbook ROI in healthcare (31% eval lift) and real estate (2.4× faster closings).

8models
16archetypes mapped

Multi-Channel Intelligence

AI agents deployed across 11 messaging channels — WhatsApp, Telegram, Slack, Discord, Teams, and more.

11channels