![⚡ Air-Gapped OCR + DSPy Vision: Build Better Agents [AI Tinkerers - Post-Training]](https://sloppy-joe-app.imgix.net/ai_tinkerers/logos/post-training_laptop-sticker_190_de2487c31e50813a3471e79627268ffb.png?w=300&fit=max&auto=format%2Ccompress&q=75){kind=small}
⚡ Air-Gapped OCR + DSPy Vision: Build Better Agents
Issue #11 · Week of December 15
Some demos from the past two weeks that popped: lots of energy around Production AI Engineering and mature Agentic Systems. Builders are moving past prototypes, focusing on reliable, integrated workflows using cheaper, powerful open-source models.
Inside, see Patrick Frenett (London) build hill-climbing environments for coding agents, and Marcus Leiwe (Hong Kong) leverage SLMs for air-gapped document intelligence. These are real, vetted, practical builds from your peers. Scan the featured builds and pull the tactic that closes your next gap.
Top 5 Picks (December 15)
1
TOP PICK
Engineer BS Detector: Authenticity Wins
Sara Miteva from PostHog presented The Engineer BS Detector: Why Authenticity Wins. The talk outlined how to build engineer-oriented products without fluff, focusing on signals engineers actually use and what they value, with AI helping only when grounded in real product data. It offered practical frameworks for tying features to observed behavior and crafting honest messaging, a timely angle as teams wrestle with authentic AI-driven value; audience feedback was positive. It’s a reminder to keep builder-focused storytelling grounded in signals.
TECH STACK
PROJECT LINKS
|
2
RUNNER UP
Dialog AI: Building Autonomous Systems
Mike Karpov from Dialog AI presented How We Build Autonomous AI Systems, showing how they design agent-like components that understand context, reason, and interact with tools to support business workflows. The talk covers architectural principles, engineering patterns, memory, and control, with a focus on production-grade stability over specific models. The Dialog AI stack powers multilingual voice agents that automate qualification, scheduling, and collections via an autodialer. Audiences responded positively to the practical angle, underscoring real-world AI orchestration for builders.
TECH STACK
PROJECT LINKS
|
3
COMMUNITY FAVORITE
GraphQL vs REST for MCPs
Piotr Zaczek from AITV presented Why I dropped using REST for MCPs and started using GraphQL. GraphQL-like interfaces for MCPs query relational data, reducing latency, token usage, and network calls. It shows a practical path to faster, cheaper agent orchestration with real-time benefits. (Audience feedback hinted strong interest.) This approach fits the current trend toward efficient, low-latency AI workflows for builders.
PROJECT LINKS
|
4
STANDOUT
Temporal Workflow Conversational Agents
Dushyant Pathak from UiPath presented a long running Temporal based conversational agent solution. The system supports multi-turn dialogues up to 12 hours, with suspension and rerun workflows, inbuilt retries, error handling, and dependency management to trigger complex workflows, and it uses a DeepRAG model to share context across storage. Key tools include Temporal workflows, DeepRAG, and UiPath Studio Web integration. Survey feedback noted practicality, with attendees appreciating robust enterprise tradeoffs. It hints at scalable, production-ready agent orchestration for enterprise automation.
|
5
NOTABLE
Podcastify AI: Spaced Repetition Lessons
Parsa Neshaei, a PhD student at EPFL, presented Podcastify AI, a watchOS app that delivers personalized, podcast-style language lessons on the go. It generates AI-driven audio lessons tailored to user level and interests, uses spaced repetition for retention, and runs on-device for low latency with Swift tooling. On-device inference and real-time adaptation meet busy learners where they are, and audience feedback suggests strong interest (people loved it). On-device GenAI can power practical, accessible learning tools for builders.
TECH STACK
PROJECT LINKS
|
More Great Builds
Quick hits from the community — demos worth bookmarking:
Marcus Leiwe • AI Tinkerers - Hong Kong • Nov 27
Marcus Leiwe from Leiwe & Partners showcased Multimodal SLMs: Qwen 2.5 and Air-Gapped Document Intelligence for Confidential Data. The demo runs on-prem in air-gapped mode using Qwen2.5-VL-7B to digitize Branches of Hope's confidential refugee records into structured data, upgrading to Qwen3-VL-8B-Instruct to compare power gains. It demonstrates OCR with multimodal LLMs, outpacing traditional tools when guided by careful prompting. This aligns with the move toward accessible, secure, local AI workflows (people loved it). It shows small models can tackle real-world docs.
Loading tech tags...
Joon Lee • AI Tinkerers - London • Nov 25
Joon Lee presented Tally, a wearable ambient AI interfacer that captures real world context with dual ear cameras and on device audio. Tally ingests realtime video via RTMP and LiveKit and builds a persistent memory of transcripts, embeddings, moments, and timelines. It includes a Speech-to-App system to turn natural language into actions and an SDK for ambient agents powered by user context. Audiences reacted positively to its hardware-to-memory path and the potential for consumer wearables with long term AI memory.
Loading tech tags...
Patrick Frenett • AI Tinkerers - London • Nov 25
Patrick Frenett from Yetmo presented Coding agents and data science: building hill-climbing environments for LLMs. He demonstrates a live setup where dozens of coding agents tackle a Spotify-style color assignment task for album covers inside a hill-climbing environment, with training data, runnable scripts, and failure views across platforms. The approach shows practical self-improvement loops, metrics, and analysis tools that builders can adopt. Audiences responded well to the tips, hinting at a productizable framework for benchmarking agents. Takeaway: thoughtful environment design plus measurable feedback accelerates agent development for builders.
Loading tech tags...
Marcos Augusto Silva Vieira Franco • AI Tinkerers - São Paulo • Nov 27
Marcos Augusto Silva Vieira Franco, a Proa student, presented DSPy + Computer Vision, showing how DSPy can pair with vision systems to improve user feedback. The project uses YOLO/ONNX vision, a DSPy LLM, and TTS to deliver real-time assistive navigation. It showcases a multi-agent setup across AI components that collaborates to interpret visuals and respond. Subtle survey feedback suggested interest in the approach, and the demo hints at practical, scalable UX improvements for intelligent interfaces.
Loading tech tags...
Vysakh Sreenivasan • AI Tinkerers - Bengaluru • Nov 29
Vysakh Sreenivasan demonstrated Diwadi's smart file browser, a desktop app with smart folder triggers that can generate files like pptx and csv using AI. The app runs offline by pulling in open-source TTS and STT models, and it provides inbuilt preview and edit for major file types. Demos showed AI-driven file creation and smart folder organization. Survey cues hinted strong interest (people loved it) for tangible offline AI tooling. It points to practical business leverage if released as a product.
Loading tech tags...
Gaurav Maheshwari • AI Tinkerers - Bengaluru • Nov 29
Gaurav Maheshwari from Oodle AI presented Production Incident (PI) to Pull Request (PR) in Minutes, a demo of AI-assisted debugging that surfaces insights from alerts. It coordinates agents with Google ADK and uses Cursor/Claude Code to go from alert to PR, with WebMCP-inspired workflows to orchestrate UI actions like creating an alert, creating Grafana dashboards via AI assistant. Survey feedback hinted strong interest, and the approach points to real-world productivity gains with potential to become a production-ready observability product; for builders, it shows how cross-tool agent orchestration speeds incident response.
Loading tech tags...
Daniel Hails • AI Tinkerers - London • Nov 25
Daniel Hails, CTO of Kular, presented Your own Cursor Tab: Reverse Engineering Tab Completion. The talk analyzes how tab completion behaves in production editors, tracing iterations from vibe-code to Cursor's behavior and highlighting context engineering, debouncing, and latency optimizations. It stood out for a practical, reproducible dive into latency reduction and real-time tool-use efficiency; survey feedback hinted strong interest from builders. If released as a product, this approach could power faster, context-aware coding assistants for developers.
Loading tech tags...
🏆 Hackathon Spotlight
Recent AI Tinkerers Hackathon Winners
The Citadel created a high-speed, neuro-symbolic Security Kernel that defeats the 90% ASR Visual Injection vector and ensures Data Sovereignty via real-time secret redaction in LLM agents. Solo hacker Johnny Hung, CEO of Nine Suns, leveraging a decade of engineering experience from compliant heavy companies like McKesson and Motive, plus a background in Visual Arts from UChicago.
🥈 2nd
MyCodeDontJiggleJiggle
Cleo is an innovative agentic workflow that uses intelligent screen capture and heuristics to automatically construct compelling narratives and marketing collateral about a developer's feature work. Built by two engineers from XBOW, including an AI Researcher specializing in offensive security agents and a veteran developer passionate about devtools.
Latest Content
How to Ship Complex Features 10x Faster with AI Agents | Dex Horthy (HumanLayer)
Dex Horthy (HumanLayer) breaks down the “12 Factor Agents” approach to shipping multi-step agentic workflows faster: structured outputs, ...
Watch Now →
How to Run Open-Source LLMs Locally on a Mac with MLX-LM
Run open-source LLMs locally on Apple Silicon with Apple’s MLX-LM: `pip install mlx-lm`, then `load()` a Hugging Face model and call `gen...
Read More →
Top Job Matches
Matched based on your meetup activity and profile
Paxos Health
• New York & Toronto
• $110k - $175k (varies w/ location/level); generous equity
Stanford-founded Seed-stage healthcare AI startup with >$5M in VC funding and AI agents deployed in production with cu...
Apply Now →
Dex
• London (5 days on-site)
• £250,000
Frontier AI engineering role building the AI tooling layer for complex financial modelling.
Apply Now →
Jakib AI
• Columbus, OH
Jakib is a profitable, growing applied AI firm embedded with operator-led companies in logistics, manufacturing, and c...
Apply Now →
You are one of 95,000+ readers from Anthropic, OpenAI, Google, Microsoft, Meta, Apple, Amazon, Nvidia, Netflix, Stripe, Databricks, Snowflake, and others — spanning frontier labs, big tech, startups, and top universities.
⚡ Air-Gapped OCR + DSPy Vision: Build Better Agents