DAVID RUSSELL

Programmer

David Russell outdoors

From the model workshop

Fine-tuning, measured against the task.

I built the path from reviewed conversations to a working local model: data preparation, adapter training, evaluation, and an application that uses the result.

Read the experiment →
1,250reviewed training exchanges
27Bparameters in the shared base
80validation and test scenarios

Selected work

All projects →
FRONTLINEExample report

Gateway inspection

COMPLETED

Reseated the power cable.

OBSERVED

Gateway back online.

UNKNOWN

Power supply rating.

Every claim has a source. Missing facts stay missing.

Live web demo

Frontline

Maintenance reports with a paper trail.

Turn notes, recordings, and photos into a maintenance report that keeps observations, completed work, and next steps distinct.

Multimodal inputs / Retrieval / Evidence review

Explore the project →
OTSCIllustrative task view
CodeExplanationChanges

Handle the empty case. Keep fractional results.

def average(values):
    if not values:
        return 0
    return sum(values) / len(values)

A useful answer. Room for the next requirement.

Mac desktop app · In development

Over The Shoulder Coder

An AI collaborator for work in progress.

Follow the screen and the conversation, keep track of changing requirements, and propose code, explanations, diffs, or designs.

Continuous context / Structured outputs / Native controls

Explore the project →
CAREER WORKBENCHFictional example

A contribution, with its context.

“I helped keep the volunteer schedule up to date. The program lead owned it.”

Supported the weekly volunteer schedule alongside the program lead.

Shared credit preserved · Original account linked

Local web app · Source available

Career Workbench

A good resume starts before the writing.

Discover overlooked experience, explore directions, and develop profiles whose claims lead back to the original evidence.

Agentic discovery / Persistent evidence / Document review

Explore the project →
QWEN TTRPGTraining + inference

One shared model

27B

Qwen base · three task adapters

ActionsStorytellingRules

Reviewed data → train → verify → compare

Local model training · Open source tooling

Qwen TTRPG

Training an AI to take its turn.

From source-bound training examples to a 27B local model: supervised adapters, preference experiments, and evaluations that distinguish better prose from better answers.

Conversational data / QLoRA + FSDP2 / Evaluation / GPU serving

Explore the project →
STORY COPILOTAuthored illustration

“I leave the hatch closed. I want to talk to the caretaker.”

CONTEXT

The player changed direction. The hatch remains closed.

PRIVATE GUIDANCE

Give the caretaker a reply. Leave the next choice to the player.

A suggestion is a possibility. The conversation is the record.

Local web app · In development

Story Copilot

Keep the story moving. Keep the players in charge.

Follow a shared story, recover the details that matter, and suggest the facilitator’s next response while preserving each character’s choices and knowledge.

Agent decisions / Conversation memory / Grounded rules

Explore the project →