| title | feature_image | aside |
|---|---|---|
Professional stuff |
false |
I have a broad skillset that I have enhanced over the years. My current focus is on data-driven solutions for complex systems. I'm especially interested in roles that aim to combine machine learning with rigorous scientific avenues. Even better if it involves a bit of Material Science!
My diverse set of experiences would make me a valuable addition to any team, and I would prove to be an asset with my quick learning abilities. I have also led collaborative multinational studies and worked on a breadth of physical computation domains.
Please feel free to reach out if you have any queries! My details are in the Contact tab.
As seems to be the case with everything now, this role covered a lot of things (and I say this positively), and has been super fun. Few things that I have been responsible for and delivered on:
- Deployed Agentic LLM systems in prod in context of a theatrical setting.
- Synthesized speech Gen AI and explored the effect of Image Gen AI in the space of performing arts.
- Evaluated feasibility of different LLMs, spanning proprietary (OpenAI, Gemini, etc) and open-source (Llama 2 + 3, Mistral, etc) models.
- Customised LLMs above using client data after identifying appropriate strategy (LoRA, ORPO, etc).
- Designed and deployed LLM tech stack in AWS Sagemaker (monitored via Shadow inferencing and Cloudwatch).
- Developed in-house software for organisation to extract data from own docs for precise LLM fine-tuning.
- Co-authored extensive multinational collaborative grants (£1m+) to address ethics in Gen AI.
This role has given me a deep insight into something I deeply care about: LLM deployment but responsibly. Ethics in Gen AI are a tricky issue and navigating this in a production environment has been extremely rewarding and it's a super awesome system!
This role covers a lot of things. Mainly, these are the few things that I have been responsible for and delivered on:
- Build up the tech stack
- Build APIs for GCP and mobile app integration
- Build MLOps pipelines
- Maintain all of the above
- Secured GCP start-up award ($350k)
All things considered, this has been a challenging and rewarding experience.
The initial challenge was to connect to the client sites and transfer data to the Pzeros database. To resolve this, I wrote some scripts on the devices that are sent out to the clients (with everything pre-installed), and the device starts uploading the images to our database as soon as it's linked up.
Once this is done, the real analysis begins. Vertex AI is great for deployment, but it's not very nice for ONNX deployments. To resolve this, I containerize the models and use the Artifact Registry (after Cloud Build fetches from GitHub) to feed into GCP as an Vetex AI endpoint.
This inference, which is obtained once any new image is uploaded, is written to BigQuery, which in turns feeds into the Looker Dashboards for the client whose images are under consideration.
This forms the core service of the firm, which you can read about more on the firm's website.
My PhD deals with high-temperature performance of materials and how they perform at a small-scale.
Broadly, this can be divided into 3 phases:
- Experimentation
- Simulation
- Machine learning
The experiments involved were a departure from the conventionally used big specimens in the industry. My research adopts miniaturized novel specimen geometries for these tests. Volumetrically, my newly designed test technique provides a materials saving of 97.89%.
My research would make it easier to inspect powerplants and jet-engines with minimal amount of material, therefore not impacting the structural integrity. This would help in the development of new alloys and test irradiated structures with a higher safety margin.
I've gone one step further in my research and I'm investigating the feasibility of combining 3 tests from a single specimen (Small Punch + Small Ring Test). Volumetrically, this resulted in materials savings of 97.89%. Combing 3 tests (creep, tensile, fatigue) is also groundbreaking for these industries. My research overview is outlined below.
{% include figure.html image="/assets/images/Website_professional_thesis.svg" caption="Research plan outline." width="800" height="1000" %}
This combination can be visualized in the image below, with the ring painted for computer-vision based strain imaging.
{% include figure.html image="/assets/images/Professional_ring_disc_combination.jpg" caption="Visualizing my hypothesis of combining ring and disc testing." width="800" height="1000" %}
I used optimization routines in conjunction with numerical simulations to obtain unknown material parameters. I leveraged this technique to translate the test's force-displacement data to a more general-purpose stress-strain data for this new testing methodology.
This data was then fed into regression models (XGBoost and Random Forests, primarily) to analyse rate-dependency of the test. The models trained well with an R2 score of 0.95632 and the predictions were extremely close on new hidden data as well, as seen in the image below.
{% include figure.html image="/assets/images/316L_RT_SRTT_0_3333mmMin_cleaned_RFRandXGB_limited.png" caption="Predictions on new experiments via RFR and XGBoost." width="800" height="1000" %}
My original experiments also performed largely well, as you can see in the heatmaps below. For the Small Ring Test (SRT) I performed 48 experiments at different pin displacement rates and for the Small Punch Test (SPT) I analysed the discs separately and both sets of inter-test comparison results are shown below:
{% include figure.html image="/assets/images/SRTT_heatmap_R2scoreMinMax.png" caption="Inter-test comparison R2-scores for all the SRT experiments, with each tick representing the pin displacement rate in mm/min." width="800" height="1000" %}
{% include figure.html image="/assets/images/SPTT_heatmap_R2ScoreMinMax.png" caption="Inter-test comparison R2-scores for all the SPT experiments, with each tick representing the pin displacement rate in mm/min in the brackets. The B prefix represents the hypothesis batch (5 batches were tested) and the disc number represented by D. The Norm represents the normally extracted discs." width="800" height="1000" %}
Extensive data cleaning procedures were also analysed, such as butterworth signal filtering (shown below) or the convolution smoothing. See the image below! I also devised a data-cleaning GUI-app for the department so that people don't have to go through the faff of creating this from scratch for their experiments. GitHub repo here.
{% include figure.html image="/assets/images/butterworth_filter_example.png" caption="Visualising the cleaning up the test input data with Butterworth filter." width="800" height="1000" %}
Using the power of physics baked in with my analytical solution, I devised a physics-informed data-driven conversion method that allows users to obtain material parameteres directly without the need of any expensive simulations for the Small Ring Test.
Below are a couple of images of the Small Ring Test as well. Images are taken from a Nikon D850 camera.
{% include figure.html image="/assets/images/SRT_initial.jpg" caption="Small Ring Test- Ring specimen (SS316L) at the start of test." width="800" height="1000" %}
{% include figure.html image="/assets/images/SRT_final.jpg" caption="Small Ring Test- Ring specimen (SS316L) at the end of test." width="800" height="1000" %}
As part of the National Transmission Project (NTP), PROTECT portfolio, I worked as a Research Assistant within the AMCG (Applied Modelling & Computation Group) at Imperial College London.
My work involved analyzing the propagation of air-flows in different built environments. My work went on to inform the knowledge domain for COVID-19 and how mitigating measures may be implemented for better health and safety of occupants in any given structure.
This work involves numerical simulations on IC-FERST for large-scale geometries/structures that are also designed with the help of an in-house software.
Unfortunately, due to the confidential nature of this work, I am not able share any cool images relating to this.
This role pertained largely to research in severe accidents in nuclear reactors, and was a part of an ongoing effort to establish international cooperation between AMCG and other academic partners. We published a hugely collaborative paper (titled 'Numerical Study of the Effect of Geometry on the Behaviour of Internally Heated Melt Pools for In-Vessel Melt Retention').
In this role, I supervised MSc projects with our partners and enhance IC-FERST's capabilities (in-house numerical simulation software) for better simulating severe accidents.
The severe accidents pertain to a cataclysmic event in the lower-head of a reactor pressure vessel, where all the fuel material collects post-accident. More specifically, this work involves simulating the cooling of this fuel debris (it emits a lot of residual heat) and analyze its safety. Below is a snapshot of the fluid dynamics in this reactor after 8 seconds of simulation.
{% include figure.html image="/assets/images/tstep8_temp_velvec_nomesh.png" caption="Velocity vectors superimposed on the temperature field background. Simulation of lower-head of a reactor pressure vessel." width="800" height="1000" %}
- Research area: Materials Science and Machine Learning.
- Thesis: Evaluation and optimisation of the Small Ring Test and its amalgamation with the Small Punch Test.
- Thesis: Numerical Modelling of In-Vessel Melt Retention: The LIVE Experiments.
- Awarded the William Penney Prize for my thesis work.
- Thesis: Analysis of shockwave propagation in air.
Aniket Joshi, Alex Forsey, Richard Moat, Salih Güngör. Physics and data driven approach to analysing the Small Ring Tensile Test on SS316L at multiple displacement rates. Under Review.
Aniket Joshi, Alex Forsey, Richard Moat, Salih Güngör. Leveraging digital image correlation for the design of a new strain measurement system for the Small Ring Tensile Test. Under Review (w/ Experimental Mechanics).
Aniket Joshi, Alex Forsey, Richard Moat, Salih Güngör. Preliminary results on feasibility of combination of the Small Ring Test and the Small Punch Test. Under Review.
Aniket Joshi, Alex Forsey, Richard Moat, Salih Güngör. Open Source Digital Image Correlation Analysis Data on Select Open Source Data for Small Ring Tensile Test. Open Research Data Online, The Open University.
Aniket Joshi, Alex Forsey, Richard Moat, Salih Güngör. "Open-Source Data for Small Ring Tensile Test performed on SS316L at multiple displacement rates." Open Research Data Online, The Open University.
Akash Venkateshwaran, Mahendhar Kumar, Shyam Kumar, R. Sivakumar, Aniket Joshi, and Christopher Pain. "Numerical study of the effect of geometry on the behaviour of internally heated melt pools for in-vessel melt retention." Progress in Nuclear Energy 156 (2023): 104555.
Aniket Joshi, Alex Forsey, Richard Moat, and Salih Güngör. "Preliminary results on conducting Small-Ring Test and Small-Punch Test on the same specimen." Transactions, SMiRT-26, 2022.
Rossella Arcucci, César Quilodrán Casas, Aniket Joshi, Laetitia Mottet, Asiri Obeysekara, Yi-Ke Guo, and Christopher Pain. "Enhancing CFD simulations of COVID-19 diffusion by coughing and sneezing using data assimilation." High Performance Computing on CRESCO Infrastructure: research activity and results 2020 (2021): 52.
Aniket Joshi, Claire Heaney, Alan Jones, Liang Yang, Paul Smith, Ali Tehrani, Christopher C Pain. "Numerical modelling of in-vessel melt retention: The LIVE Experiments." 25th International QUENCH Workshop: 22-24 October 2019, Karlsruhe Institute of Technology, Karlsruhe, Germany.