Selected work
A domain-adapted investment research model
NYIC leadership experience adapting Qwen2.5-14B-Instruct for investment research and deploying it on dedicated GPU infrastructure.
Engineering lessons
What connects a domain-adapted model to a useful production service?
- The study connects a defined Qwen-based adaptation program with dedicated serving infrastructure.
- Dataset lineage, task behavior, and inference configuration belong in the same release record.
- Interpret performance only with its task composition, baseline, and serving conditions.
Context and professional contribution
Investment research assistants must handle specialized language, structured outputs, tool interactions, and appropriate limits on unsupported conclusions. A general model can provide a strong starting point, while a domain-adaptation program can target stable behaviors that recur across the research workflow.
This study describes NYIC leadership experience implementing an open-weight investment research copilot. The source portfolio records fine-tuning of Qwen2.5-14B-Instruct with QLoRA and parameter-efficient fine-tuning methods, followed by deployment through vLLM and TensorRT-LLM on NVIDIA DGX-backed Kubernetes infrastructure.
The project connected a defined adaptation objective with a deployable model artifact and a dedicated serving environment. This made dataset design, task behavior, and inference configuration part of the same engineering problem.
Adapt a defined model to domain tasks
The documented training corpus contained approximately 38,000 deidentified examples spanning research, filings, earnings calls, tool use, and refusal behavior. This mixture connected domain material with the actions and limits required of the assistant. The work adapted an existing 14-billion-parameter instruction model to the research task and its expected interaction patterns.
The inclusion of tool-use examples is significant because a research assistant often needs to obtain current facts or invoke analytical services. Demonstrations can teach the relationship between a task, the tool arguments it requires, and the result that should be incorporated into the response. Refusal examples can represent conditions in which the model should decline or request clarification.
For this type of dataset, deidentification is an important processing step with limits. Information ownership, permitted use, and residual identifying context still require review. Those concerns should remain part of the model lifecycle even when direct names or account identifiers have been removed.
Connect adaptation to the inference environment
QLoRA and parameter-efficient fine-tuning made the adaptation program compatible with a constrained trainable parameter footprint. The resulting model was deployed with vLLM and TensorRT-LLM on DGX-backed Kubernetes. These serving choices connected the model artifact to an operational inference environment.
Training and serving need to be assessed together. Prompt formatting, context handling, quantization, and execution configuration can affect the behavior observed by a consuming application. A model release therefore needs a reproducible configuration that connects the checkpoint or adapter to the interface used in production.
The project description also sits within broader leadership experience building large-scale model-serving infrastructure. That experience includes multi-GPU and multi-node execution, runtime profiling, workload isolation, release controls, and rollback. Those platform capabilities are relevant context, while this case’s specific model deployment remains the Qwen-based research copilot.
Understand the serving tradeoffs
A dedicated open-weight deployment gives the organization direct control over the model artifact and serving configuration. It also introduces responsibility for capacity, upgrades, runtime behavior, and operational support. The appropriate comparison includes the full service lifecycle as well as inference charges.
Request shape matters. A short tool-selection task and a long filing-analysis request can consume very different amounts of context memory and compute. Capacity planning should therefore use a representative workload distribution and inspect tail latency, throughput, and recovery behavior under concurrency.
The source portfolio records deployment through two inference engines. A useful engineering practice is to keep a stable application contract while validating each engine’s behavior and configuration. This makes performance experiments interpretable and reduces the chance that an interface change is confused with a model-quality change.
Evaluate domain behavior with explicit context
A useful performance comparison requires the full task composition, comparison-model versions, traffic distribution, and cost boundary. Those conditions determine how a result should be interpreted and whether it is relevant to a new enterprise workload. Task quality and serving economics need to be measured together.
A reproducible assessment should retain the task definition, evaluation examples, baseline model identities, prompt configuration, serving conditions, and scoring rubric. It should separate successful formatting from correct research behavior and examine whether tool use and refusal remain appropriate on unseen cases.
Domain adaptation also needs tests for overfitting and unintended capability changes. Useful cases include unfamiliar issuers, revised document structures, contradictory evidence, missing tool results, and requests outside the intended authority. Evaluation should compare the adapted system with a credible baseline using the same task interface.
Operating controls and model lifecycle
The broader source portfolio documents governance work covering risk classification, lineage, evaluation, access management, oversight, monitoring, incident response, and rollback. It also records model cards, risk assessments, evaluation reports, approval records, and change histories. These artifacts provide a useful operating context for adapted-model programs.
For a new deployment, feedback should be reviewed before becoming training material. A user’s acceptance of an answer can indicate usefulness without establishing correctness. Verified corrections and representative failures can improve evaluation coverage and guide a subsequent adaptation cycle.
A model release remains one component in the research application. Retrieval, financial tools, user permissions, and the review interface continue to determine whether the overall system produces an acceptable work product.
Documented result and relevance
The documented outcome is a domain-adapted Qwen2.5-14B-Instruct research model, trained on approximately 38,000 deidentified examples and deployed on dedicated GPU infrastructure. The experience connects dataset construction, adaptation, and serving within one implementation path.
The central lesson is to judge adaptation through the task and the operating environment together. Model identity, data provenance, evaluation conditions, and serving configuration provide the context needed to decide whether a specialized model is the right component for an enterprise workflow.