About
I am an NSF NRT Fellow at the University of Delaware's Center for Composite Materials, advised by Dr. Srikanth Pilla, and part of AIM for Composites, an Energy Frontier Research Center funded by the U.S. Department of Energy. I passed my Ph.D. preliminary exam in August 2026 and expect to graduate in May 2028.
My work sits between machine learning and engineering. One project looks at a photo of a car and works out how each part was manufactured. The other is a service that has read more than ten thousand open-access papers and built a database of 2.27 million material property values from them, with checks that test those values against each other. In both, most of my effort goes into finding out where the model is wrong: benchmarks, preregistered tests, held-out data and audits against human labels.
Before the Ph.D. I was an HPC performance engineer at Cornelis Networks. I hold a B.S. in Mathematics and Economics from the University of Delaware (2024).
Research
2026 to present
Paper in preparation
An autonomously built composites database, and how to check it
Property data for fiber-reinforced composites sits mostly in research papers: in tables, in sentences and in plots. I built a service that reads the open-access literature and builds the database from it. It is one program on a scheduler, deployed on a Hugging Face Space and writing to PostgreSQL. No person selects, supplies or corrects a query, a document or a value.
One cycle of the service
- Discover. A discovery pass fills a queue of candidates from OpenAlex, Europe PMC, Semantic Scholar, arXiv, NASA's technical report server and vendor data-sheet pages. The queries come from a rotating set of seeds, a grid of fiber, matrix and property terms, and a language model.
- Download. Only open-access copies are read. Duplicates are removed by DOI, link and file hash, and a file is accepted only if it verifies as a PDF.
- Extract. Gemini reads the text of every page and returns each value as a row: material, grade, matrix, fiber, fiber fraction, property, value as printed, unit, test condition, page and quotation. Code then lists the numbers on each page that no row used, and a second request asks only about those. Values read off figures are stored as estimates.
- Check against the source. No model is involved. Each value is searched for on its cited page, and units are converted to SI and tested against wide ranges. A row that passes is called verified. That means it passed these checks, not that it is known to be correct.
- Store. A row that fails is stored with its reason and is never deleted.
Snapshot of 30 Oct 2026
What one campaign produced
From 18 to 30 October 2026 the service added 10,024 documents in 11.1 days, to 696 it already held. It averaged 899 documents a day and read 3,028 on its best day. Extraction cost 4.1 US cents per paper at Gemini's list price.
| In the database | Count |
|---|---|
| Documents | 10,720 |
| Property values | 2,270,377 |
| Values read from text | 1,734,313 |
| Of those, verified against the source | 1,465,233 (84.5%) |
| Values read off figures, kept as estimates | 536,064 |
| Values for polymers, composites and fibers | 1,374,954, 625,821 and 269,602 |
6,181 hand-labeled values in 52 papers, scored 10 Nov 2026. Matching is automatic.
How much of a paper it finds
The database held 81.6% of the labeled values (95% confidence interval 76.9 to 86.1%), and at least 74.0% once chance matches are subtracted. For a property in its vocabulary that a paper states in its own text or tables, outside long series, it held 95.5%. The single-pass extractor it replaced held 36.2% on the same PDFs.
This measures coverage, not correctness. The annotators recorded a median of 113 values per paper, and the database holds 231.
The idea of the paper
Checks that borrow their inputs
A check against the source has a blind spot. It passes any error that the source itself contains: a unit printed wrongly in the paper, or a number quoted from another specimen's row. To see those, a check has to compare the value with something outside the passage.
Composite mechanics offers such comparisons, because a composite's stiffness, strength and density are tied to those of its fiber and matrix. But a paper rarely reports every input. Only 2.6% of the composite moduli in the database come with their own paper's fiber and matrix moduli.
So the checks borrow. A knowledge graph, rebuilt from the database by rules alone, links each value to its material, fiber, matrix and grade. A check takes the inputs a paper lacks from other papers' verified records of the same constituents, and every flag cites the rows it used. There are four physical relations (a rule-of-mixtures bound on modulus, a density rule of mixtures, a strength cap, and the identity between weight and volume fraction) and a consensus check across papers. Borrowing lets each relation test 4.4 to 11 times as many values.
31,638 planted errors whose source agrees with the wrong value. Results of the draft of 16 Nov 2026.
What they catch
| Method | Planted errors caught |
|---|---|
| Searching for the value in its source | 1% |
| A language model re-reading the quotation | 12 to 15% |
| Modulus bound, inputs from the same paper only | 7% |
| Modulus bound with borrowed inputs | 46% |
| Consensus across papers | 57% |
| A language model judging from its own knowledge | 65 to 72% |
The bounds are scored on composite tensile moduli and consensus on values compared across papers. The two language-model rows give both.
The language-model judge catches the most, but it rejects 10.5% of correct values and cites no evidence. Every flag from the graph names the papers and the rows it used.
Manufacturers' data sheets give an outside reference. Of 602 fiber values for named Toray and Hexcel grades, from 193 papers, 80% are within 10% of the manufacturer's number. Of the values the graph flags, 65% are more than 30% off, against 7% of the rest.
A blind human audit of 260 stored values is under way to measure correctness directly.
Manufacturing Intelligence Pipeline
Choosing how to make a polymer-composite part (injection molding, compression molding, thermoforming, filament winding and 13 other processes) is an expert decision governed by constraints. The pipeline makes that decision from a photo and shows its reasons. A vision-language model (VLM) reads the part. A rule engine, written from manufacturing textbooks and papers, turns what the model read into a ranking of the processes and lists the rules that moved each one up or down.
From photo to answer
- Detect and segment every component in a photo of a vehicle.
- Perceive. The VLM answers a fixed list of questions about each part: is it hollow, does it have ribs, are there undercuts, what is the production volume. It also gives its own shortlist of eight processes.
- Decide. The rule engine combines the answers and the shortlist into a ranking of the 17 processes.
- Explain. Each result lists the observations and the cited rules behind it, so a person can check the reasoning.
100 held-out photos, scored once in Oct 2026. Every setting was chosen on 535 development photos and frozen first. Labels are by part type; labels for each photo are in progress.
Where it stands
The benchmark is 635 openly licensed photos of real parts in 41 part categories. A part usually has about two correct processes. On the held-out photos the current design ranks a correct process first for 73% of photos and keeps every correct process in its top 8 for 88%. The first design reached 19% and 11%.
| Method | Top pick correct | All correct in top 3 | All correct in top 8 |
|---|---|---|---|
| Guessing the most common processes, without the photo | 58% | 28% | 56% |
| First design: 92 rules as vetoes | 19% | 5% | 11% |
| VLM alone | 74% | 47% | 88% |
| Rules as evidence, all 564, with the VLM | 73% | 48% | 88% |
| Rules as evidence, the 142 checked on photos, with the VLM | 72% | 45% | 90% |
| The 142 checked rules alone, without the VLM's shortlist | 36% | 39% | 70% |
In the first design every rule was a veto, so one misread feature deleted the right answer. On 11 of 16 wound composite tubes the VLM reported ribs, probably the winding bands, and the no-ribs rules removed both correct processes. I rewrote the engine so that rules are evidence: each one moves a process up or down the ranking, and nothing is deleted. The rule base grew from 92 to 564 cited rules. I checked every rule I could against the development photos, and 142 point the right way.
The rule engine is now level with the VLM on the top pick, and it adds what the VLM cannot give: the rules behind every ranking. With 100 photos, differences of a few points are within noise.
Cross-validation on the 535 development photos, Oct 2026. Every photo is scored by a model that never saw it. The held-out photos have not been scored with this model yet.
Learning how far to trust each observation
Hand-set weights treat every observation as equally reliable, and they are not. The VLM's guess of production volume is useful. Its reading of wall thickness points the wrong way. So I let the development photos set one trust number for each perceived feature, about 22 learned numbers in all, while every rule keeps the direction it was written with.
| Model | Top pick correct | All correct in top 3 | All correct in top 8 |
|---|---|---|---|
| VLM alone | 77.0% | 55.1% | 88.2% |
| Learned trust per feature, new photos | 77.6% | 57.2% | 91.2% |
| Learned trust per feature, part types never seen in training | 77.4% | 57.0% | 90.8% |
This version is ahead of the VLM alone at keeping every correct process in the top 8: by 3.0 points on new photos (95% range 1.1 to 4.9) and by 2.6 points on part types it never saw (0.9 to 4.5). The top pick does not change. Giving every rule its own weight looked better on familiar parts and worse on unseen ones, which is memorizing part types, so I froze the simpler model.
SmolVLM-256M with LoRA, trained on a CPU, Oct 2026. The same 100 held-out photos, scored once.
A small open model, fine-tuned
I also fine-tuned a small open model on the development photos. On the held-out photos every correct process is in its top 8 for 73% of photos, against 40% for the same model without fine-tuning, and in its top 3 for 43% against 29%. Its top pick did not improve: it answers injection molding for 88 of the 100 photos. On part types it never saw, it is no better than guessing, and the much larger Gemini model still leads at 88%. Fine-tuning taught the small model which processes go together. It did not teach it to see them.
Simulation, Jul 2026. 150 photos, then 635 with a preregistered protocol, Aug 2026. Audit of the metric, Sep 2026.
How the design got here
I defended the first design at my preliminary exam. On a seeded benchmark of 1,000 instances, built from 14 expert-labeled components with simulated perception noise, it scored F1 0.385 on the hard cases, where injection molding is not the only answer. A simulated direct classifier with the model's bias scored 0.136, and the rules with perfect perception 0.510.
On real photos the result inverted. On the first 150 photos the vetoes collapsed to 0.086, and the record of observations showed where: 84% of the wrong eliminations came from two yes/no questions, whether the part is hollow (58%) and whether it needs continuous fiber (26%). I reworded the hollow question and preregistered the rerun on 635 photos, with the line that would count as success written down first. The pipeline rose to 0.198 (95% confidence interval 0.176 to 0.223). That is a real improvement, and it is under the 0.20 line I had set, so I reported it as a miss and changed the design.
Then I audited the metric. A predictor that always answers injection molding scores F1 0.555 on all 635 photos, so the number had no floor. I replaced it with the plain percentages above, always next to a guessing baseline.
The same runs show the model's bias. Asked for one answer, the VLM named injection molding for 71% of 150 photos, where surveyed experts chose it in 24% of their picks. Its shortlist of eight still held every correct process for 133 of the 150.
In progress
What comes next
More rules stopped helping because the questions were the weak point. The VLM said “continuous fiber” for 86% of injection-molded parts, which a photo cannot show, and when asked twice whether a part is hollow it changed its answer on 47% of photos.
The follow-up asks about marks a camera can see: a pinch-off seam means blow molding, layer lines mean 3D printing, ejector-pin marks mean injection molding. I set the bar before collecting any data: every correct process in the top 3 on at least 52% of held-out photos, against 47% for the VLM alone. Labels for each photo, instead of each part type, are being added to the held-out set.
Beyond this, I am starting to extend the pipeline to inverse design: inferring the manufacturing process from a material's microstructure.
First-author poster, 2025
Can an LLM write a manufacturing dataset?
I generated a manufacturing-process dataset with an LLM and validated it against a survey of domain experts. The model turned out to be a knowledgeable generalist and an unreliable specialist, with gaps that follow the type of component.
M. Heider, A. M. Deshpande, S. A. Pradeep, S. Pilla. Generation of Manufacturing Processes Dataset Using Large Language Models. AIM for Composites EFRC, 2025.
Ongoing
InDeS
I lead the design of InDeS, the center's AI-enabled inverse-design software platform, and I maintain the center's GitHub and Hugging Face organizations.
Projects
Nov 2026 to present
Team project, private repository
Verification for a game simulator
A deterministic eight-player game simulator with a Python game core and a TypeScript combat engine. My part is showing that it is right.
- I built the verification harness: a benchmark that replays recorded real fights, an auditor that recomputes every hit and attack timer from the rules, invariant, mirror and mutation checks, and CI.
- The engine now calls the winner of 79.5% of 1,261 real fights, against 72.4% for a board-value baseline. Changes are gated on a paired no-regression test: exact McNemar on winner accuracy and a bootstrap on Brier score.
- I set up and run the agent loop that drives the work. It runs the checks, files a ticket for each root cause, fixes tickets in parallel waves of coding agents, and gates each pull request on the checks. I review and merge what passes.
- 32 pull requests so far, in a codebase of about 50,000 lines of Python and 35,000 lines of TypeScript with more than 1,400 tests.
Spring 2023
Parallel programming in C
Coursework: image convolution with OpenMP and, with a team, an n-body simulation in CUDA. Image convolution code
2022
2021
Breast cancer prediction
Machine-learning course project on diagnosing breast cancer. Slides
Experience
Jun 2023 to Mar 2024
HPC Performance Engineer at Cornelis Networks
- Optimized Microsoft and Intel machine-learning and computer-vision models for the Omni-Path Architecture: 42% better performance scalability and 24% faster training on larger distributed workloads.
- Adapted and parallelized AI/ML workloads for CPU/GPU-accelerated hardware, cutting test time by 18%, and validated execution on the Omni-Path fabric.
May 2022 to Feb 2023
NSF REU Researcher at Emory University
Hamiltonian-inspired neural networks, with Dr. Deepanshu Verma and Dr. Lars Ruthotto. Instead of learning a system's motion directly, the network learns its Hamiltonian, the energy function, and the motion follows from it. I built the networks in PyTorch and compared them with residual networks on spring-mass and two-body dynamics. With symplectic Verlet steps the network reached accuracy comparable to fourth-order Runge-Kutta in about a quarter of the training time: 41.6 seconds against 156.8 for 500 epochs on the two-body problem. I presented the work at the Joint Mathematics Meetings in 2023. Code
Aug 2021 to Jul 2023
Undergraduate Researcher at the Computational Research and Programming Lab, University of Delaware
Senior thesis with Dr. Sunita Chandrasekaran: predicting the vital status of acute myeloid leukemia patients from TARGET clinical and gene-expression data.
- Matched 716 samples across the two sources and reduced 58,451 gene features to 17,817.
- Compared three feature-selection methods (SHAP, PCA, differential expression) across three classifiers (random forest, gradient boosting, LightGBM), with 5-fold cross-validation, a held-out test set, a random-feature baseline, and balanced against unbalanced training. SHAP with a random forest was strongest.
- AUC was misleading on this imbalanced cohort, so I reported sensitivity and specificity with it.
- Moved training to the lab's pipeline on the DARWIN HPC cluster, which cut run time 6 to 7 times.
2021
NSF REU Researcher at Duke University
Meeting the Grand Challenges in Engineering Fellow in the Brinson Lab. I built an automated pipeline that curates polymer-nanocomposite data for NanoMine, a materials data resource under the Materials Genome Initiative. Slides
2020
Undergraduate Researcher at the Center for Composite Materials, University of Delaware
Computational methods to raise throughput and material quality on the TuFF composite line.
Aug 2021 to Jun 2023
Teaching Assistant, Data Structures at the University of Delaware
Ran a lab section, held office hours and graded.
Papers and talks
In preparation
Borrowed-input checks: validating an autonomously built composites database across ten thousand papers. Preprint expected late 2026.
In preparation
Manufacturing Intelligence Pipeline: vision-language perception and symbolic rules for manufacturing-process identification.
Poster, 2025
M. Heider, A. M. Deshpande, S. A. Pradeep, S. Pilla. Generation of Manufacturing Processes Dataset Using Large Language Models. AIM for Composites EFRC.
Talk, 2023
Predicting Dynamics of ODEs using Residual and Hamiltonian-Inspired Neural Networks. Joint Mathematics Meetings.
Skills and awards
Languages
Python, C++, C, SQL, TypeScript, MATLAB
Machine learning
PyTorch, Hugging Face, scikit-learn, XGBoost and LightGBM, SHAP, NumPy and pandas; VLM fine-tuning; LLM agents (Gemini API, agentic coding loops); benchmark and evaluation design
Systems
MPI, OpenMP, CUDA, HPC clusters, AWS, PostgreSQL, SQLite, Git, GitHub Actions CI, pytest
Awards
- NSF NRT Fellowship
- University of Delaware Academic Enrichment Award (research funding)
- Second place, Emory Mathematics Department research competition
- Meeting the Grand Challenges Fellow (Duke and NSF)
Outside the lab
I compete in Teamfight Tactics. In 2026 my team finished in the top 8 of the Esports World Cup (4v4), where I was both a player and the team manager. I also founded a paid coaching business and run it end to end: curricula, pricing, client acquisition and the websites.
Otherwise: basketball, hiking and books.