Applied ML · Production Systems

Autonomous AI agents,
deployed at scale.

Specializing in autonomous AI agents, multi-agent LLM simulations, and secure, scalable AI infrastructure. I build systems that bridge the gap between complex data science and production-grade engineering.

View case studies Try the playground →
Currently
Senior Applied ML Engineer
Autonomous AI · LLM Infra
Stack
Python · LangChain · FastAPI
Local LLMs · Vector Databases
Experience
Multi-Agent Systems
5 core systems shipped
pipeline://dewi-ai-platform/live-stream
throughput 1,247 req/s p99 47ms uptime 99.94%
01 · INGEST AUTH & API fastapi · postgres 2.1k rps 02 · VALIDATE SECURE FILES multimodal parsing 0.04% fail 03 · ROUTING DYNAMIC model selection 3.4k qps 04 · INFER ZHIPU GLM real-time reasoning 14m wall 05 · MEMORY PERSIST user context 12 versions 06 · LIVE API chatdewi 47ms p99
5
core systems deployed
144
autonomous personas
~80%
research cost reduction
100%
private infrastructure
Currently shipping live · 2024.11.18 14:32 UTC
dewi-ai-platform-v1
status live
p99 47ms
uptime 99.94%
traffic 100%
synthetic-communities-v2
status live
agents 144
markets 6
traffic 100%
data-science-agent-v1
status batch
cadence on-demand
stack pandas/numpy
turnaround mins
geo-retrieval-framework-v1
status canary
llms 3
uptime 99.91%
traffic 10%
Case Studies
§ 02 / 06

Five systems, deployed and measuring.

Each card opens to a side-by-side of the business problem and the deployed architecture, the full tech stack, and the metrics that survived contact with real traffic. No vanity benchmarks — these are production numbers, measured at the gateway.

01

Dewi AI Platform

fastapi · postgresql · zhipu glm api
Enterprise-grade frontier AI assistant and document analysis platform built from scratch on private infrastructure.
100%
Private Infra
+
Business problem

Launched a secure, enterprise-grade frontier AI assistant and document analysis platform built entirely from scratch on private infrastructure. By implementing dynamic model routing and multimodal capabilities, the platform delivers real-time, reasoning-based AI responses while optimizing for speed and cost.

It’s a fully scalable solution featuring user authentication, persistent memory, and secure file processing delivering complete control over data and architecture.

View Platform -> Frontend designed and coded by Haley Garner
Tech stack
python fastapi postgresql zhipu glm dynamic routing multimodal auth
Deployed architecture
FASTAPI gateway AUTH validation ROUTER dynamic POSTGRES memory SECURE FS files GLM API inference CLIENT RESPONSE TELEMETRY
02

Synthetic Communities Engine

python · local llm (glm-4) · multi-agent
Multi-agent simulation mimicking real-world information flow across 6 global markets. Automated creation of 144 autonomous personas.
144
Autonomous Personas
+
Business problem

Engineered a multi-agent simulation mimicking real-world information flow across 6 global markets. Automated the creation of 144 autonomous personas with distinct psychographics and network influence mapping.

Acts as a predictive focus group, simulating reaction dynamics to identify viral vectors and blockers before launch, reducing traditional market research costs by ~80%.

Tech stack
python local llm glm-4 multi-agent psychographics network mapping
Deployed architecture
MARKETS 6 global PERSONAS 144 agents NETWORK influence map PSYCHO profiles GLM-4 local llm SIM engine REPORT VIRALITY METRICS
03

GEO & Brand LLM Retrieval

gpt · claude · gemini
Quantitative framework to measure Generative Engine Optimization (GEO) effectiveness and narrative positioning across multiple LLMs.
3
LLMs Analyzed
+
Business problem

Developed a quantitative framework to measure Generative Engine Optimization (GEO) effectiveness, analyzing mention rates and narrative positioning across GPT, Claude, and Gemini to optimize brand visibility.

Tech stack
python gpt claude gemini quantitative
Deployed architecture
BRAND context PROMPTS templated LLMS gpt/claude/gem PARSER mentions NARRATIVE positioning SCORER geo metric REPORT VISIBILITY DASHBOARD
04

Data Science AI Agent

langchain-style · pandas · numpy · matplotlib
Autonomous "Data Strategist" agent that chains LLM reasoning with statistical rigor for end-to-end workflow automation.
Days→Mins
Turnaround Time
+
Business problem

Architected an autonomous "Data Strategist" agent that chains LLM reasoning with Pandas x NumPy for statistical rigor and Matplotlib for visualization.

Automates the end-to-end data science workflow—from raw ingestion to insight generation—reducing analysis turnaround time from days to minutes.

Tech stack
python langchain-style pandas numpy matplotlib autonomous
Deployed architecture
RAW DATA ingest LLM reasoning AGENT strategist PANDAS transform NUMPY statistics MPL visualize INSIGHTS CHARTS SUMMARY
05

Secure Infrastructure Dashboard

python · flask · nginx · cloudflare tunnel
Real-time web-based monitoring dashboard on a resource-constrained Raspberry Pi. Aggregates hardware telemetry and local LLM status securely.
1
Raspberry Pi
+
Business problem

Designed and deployed a secure, real-time web-based monitoring dashboard on a resource-constrained Raspberry Pi. The system aggregates live hardware telemetry, authenticated Pi-hole v6 API data, and local LLM status, all securely exposed to the internet via a Cloudflare Tunnel without opening local router ports.

View Live Dashboard ->
Tech stack
python flask nginx cloudflare tunnel raspberry pi pi-hole api
Deployed architecture
CLOUDFLARE tunnel NGINX reverse proxy FLASK dashboard PI-HOLE v6 api LOCAL LLM status HW telemetry UI RPI SECURE
MLOps · Deployment topology
§ 03 / 06

From commit to inference, end-to-end.

The full topology behind the case studies. Hover any node to see its role and throughput. Drag to pan, scroll to zoom — the mesh is dense on purpose, because production is dense.

topology://davidisaacson.dev/production
01 · INGEST SOURCES api · web · postgres 02 · QUALITY VALIDATE pydantic · auth 03 · FEATURES VECTOR DB embeddings · rag 04 · TRAIN AGENTS orchestrator 05 · REGISTRY MLFLOW v4.2.1 · prod 06 · STREAM CLOUDFLARE tunnel 07 · ONLINE REDIS memory cache 08 · TRACKING EXPERIMENTS 3,847 runs 09 · SERVE TRITON tensorrt · live 10 · EDGE GATEWAY fastapi · live 11 · OBSERVE PROMETHEUS
throughput
Experiments · Ablations
§ 04 / 06

What worked, what didn't, measured.

Loss and accuracy curves from the Synthetic Communities model's training run, latency benchmarks across LLM variants, and an ablation table showing the marginal contribution of each agent feature group. The numbers below are pulled directly from the MLflow tracking server.

Loss · synthetic-communities-v2 · 100 epochs
train validation
F1 score · validation set
val F1
Latency benchmarks · p99 across LLM variants
measured at gateway · 1k req/s load
local-glm-4-cpu
220ms
zhipu-api-call
410ms
rag-pipeline
540ms
data-agent-exec
890ms
multi-agent-ensemble
1.2s
Ablation study · synthetic-communities-v2 5 variants · val set n=9,400
Variant F1 AUC-ROC LogLoss p99 latency Δ F1 vs baseline
full model (v2.0) 0.847 0.912 0.318 1.2s —
− persona memory 0.821 0.891 0.352 1.1s −0.026
− psychographics 0.798 0.864 0.391 1.0s −0.049
− network influence mapping 0.812 0.878 0.372 1.2s −0.035
− multi-agent debate 0.847 0.901 0.344 0.8s 0.000
Confusion matrix · synthetic-communities-v2 · validation set
n=9,400 · threshold=0.5 · calibrated
Predicted Viral Predicted Block Total
Actual Viral 847 153 1,000
Actual Block 92 8,308 8,400
Total 939 8,461 9,400
Precision
0.902
Recall
0.847
F1
0.874
Playground · Live API tester
§ 05 / 06

Send a payload. Get a prediction.

This is the same endpoint powering the case studies, sandboxed. Edit the JSON payload, hit send, and watch the request travel through the gateway to the inference server and back. Latency is measured at the client, not the server.

POST
api.davidisaacson.dev
Request body application/json · UTF-8
press ⌘↵ to send · latency measured at client
response idle
— ms
Response body awaiting request
// no request sent yet // click "Send Request" or hit ⌘↵ // response will appear here with a latency badge
Headers collapsed
content-type: application/json x-request-id: req_— x-model-version: — x-inference-region: us-east-1